Fix PnP rotation/translation shape for modern OpenCV/NumPy.

solvePnP returns (3,1); transpose made (1,3), which broke eye depth .dot()
and struct.pack of translation scalars. Flatten to (3,) instead.

Co-authored-by: Cursor <cursoragent@cursor.com>

.github/actions/upload-dated-artifact/action.yaml

@@ -0,0 +1,26 @@
+name: Upload dated artifact
+description: Uploads an artifact with a datetime
+
+inputs:
+  name:
+    description: The name of the artifact
+    required: true
+  path:
+    description: The path to the file/folder to upload
+    required: true
+
+runs:
+  using: composite
+  steps:
+    # https://stackoverflow.com/questions/60942067/get-current-date-and-time-in-github-workflows
+    - name: Get current datetime
+      id: date
+      shell: bash
+      run: echo "::set-output name=date::$(date '+%Y-%m-%d_%H-%M-%S')"
+
+    - name: Upload artifact
+      uses: actions/upload-artifact@v3
+      with:
+        name: ${{ inputs.name }}_${{ steps.date.outputs.date }}
+        path: ${{ inputs.path }}
+        if-no-files-found: error

.github/workflows/create-binaries.yaml

@@ -0,0 +1,87 @@
+name: Package using PyInstaller
+on:
+  push:
+    tags:
+      - v**
+  workflow_dispatch:
+
+jobs:
+  build-windows:
+    env:
+      dist-path: dist/OpenSeeFace
+    runs-on: windows-latest
+    steps:
+      - name: Pull OpenSeeFace
+        uses: actions/checkout@v3
+
+      - name: Create binary
+        shell: cmd
+        run: |
+          choco install vcredist140
+
+          xcopy c:\Windows\System32\msvcp140.dll .
+          xcopy c:\Windows\System32\vcomp140.dll .
+          xcopy c:\Windows\System32\concrt140.dll .
+          xcopy c:\Windows\System32\vccorlib140.dll .
+
+          make_exe.bat
+
+      - name: Move files for uploading
+        shell: bash
+        run: |
+          # Prevents the parent dist/ directory from being included
+          mv dist/facetracker ${{ env.dist-path }}
+          cp -r models ${{ env.dist-path }}
+
+      - name: Upload artifact
+        uses: ./.github/actions/upload-dated-artifact
+        with:
+          name: windows
+          path: dist # Add the parent directory to prevent a tarbomb
+
+  # Files must be packed into a tar in order for file permissions to be preserved
+  build-unix:
+    env:
+      artifact-path: OpenSeeFace
+      tar-path: OpenSeeFace.tar
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: linux
+            os: ubuntu-latest
+          - name: mac
+            os: macos-latest
+    runs-on: ${{ matrix.os }}
+    steps:
+      - name: Pull OpenSeeFace
+        uses: actions/checkout@v3
+
+      - name: Create binary
+        shell: bash
+        run: |
+          chmod +x make_exe.sh
+          ./make_exe.sh
+
+      - name: Move files for uploading
+        shell: bash
+        run: |
+          # Prevents the parent dist/ directory from being included
+          mv dist/facetracker ${{ env.artifact-path }}
+          cp -r models ${{ env.artifact-path }}
+
+      - name: Fix permissions from PyInstaller
+        shell: bash
+        run: |
+          find ${{ env.artifact-path }} -type f -exec chmod -x {} +
+          chmod +x ${{ env.artifact-path }}/facetracker
+
+      - name: Pack files into tar
+        shell: bash
+        run: tar -cvf ${{ env.tar-path }} ${{ env.artifact-path }}
+
+      - name: Upload artifact
+        uses: ./.github/actions/upload-dated-artifact
+        with:
+          name: ${{ matrix.name }}
+          path: ${{ env.tar-path }}

Examples/Animations/JawOpen.anim.meta

@@ -0,0 +1,8 @@
+fileFormatVersion: 2
+guid: 697b139e69dcb414498dcd93210620ff
+NativeFormatImporter:
+  externalObjects: {}
+  mainObjectFileID: 0
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Examples/OpenSeeWebcamInfo.cs

@@ -0,0 +1,270 @@
+using System;
+using System.Diagnostics;
+using System.IO;
+using System.Net;
+using System.Runtime.InteropServices;
+using System.Text;
+using System.Collections;
+using System.Collections.Generic;
+using UnityEngine;
+
+namespace OpenSee {
+
+[Serializable]
+public enum OpenSeeWebcamType { Unknown = -1, DirectShow, Blackmagic };
+
+[Serializable]
+public enum OpenSeeWebcamFormat { Any = 0, Unknown = 1, ARGB = 100, XRGB, I420 = 200, NV12, YV12, Y800, YVYU = 300, YUY2, UYVY, HDYC, MJPEG = 400, H264 };
+
+[Serializable]
+public struct OpenSeeWebcamCapability {
+    public int id;
+    
+    // DirectShow
+    public int minCX;
+    public int minCY;
+    public int maxCX;
+    public int maxCY;
+    public int granularityCX;
+    public int granularityCY;
+    public int minInterval;
+    public int maxInterval;
+    public int rating;
+    public OpenSeeWebcamFormat format;
+    
+    // Blackmagic
+    public int bmTimescale;
+    public int bmFrameduration;
+    public int bmModecode;
+}
+
+[Serializable]
+public class OpenSeeWebcam {
+    public OpenSeeWebcamType type = OpenSeeWebcamType.Unknown;
+    public int id;
+    public string name;
+
+    // DirectShow
+    public string path;
+    
+    // Blackmagic
+    public long bmId;
+    public int bmFlags;
+
+    // General
+    public OpenSeeWebcamCapability[] caps;
+    
+    public List<string> prettyCaps = null;
+    
+    private List<Tuple<OpenSeeWebcamCapability, int>> splitCaps = null;
+    
+    bool ge(bool a, bool b) {
+        return a || (!a && !b);
+    }
+
+    private int CompareCaps(OpenSeeWebcamCapability a, OpenSeeWebcamCapability b) {
+        int aRes = a.minCX * a.minCY;
+        int bRes = b.minCX * b.minCY;
+        bool aGoodFps = a.minInterval <= 670000;
+        bool bGoodFps = a.minInterval <= 670000;
+        bool aGoodRes = aRes >= 850000 && aRes <= 2200000;
+        bool bGoodRes = bRes >= 850000 && bRes <= 2200000;
+        if (a.rating < b.rating && ge(aGoodFps, bGoodFps) && ge(aGoodRes, bGoodRes))
+            return -1;
+        if (a.rating > b.rating && ge(bGoodFps, aGoodFps) && ge(bGoodRes, aGoodRes))
+            return 1;
+        if (aRes > bRes && ge(aGoodFps, bGoodFps) && ge(aGoodRes, bGoodRes))
+            return -1;
+        if (aRes < bRes && ge(bGoodFps, aGoodFps) && ge(bGoodRes, aGoodRes))
+            return 1;
+        if (a.minInterval < b.minInterval)
+            return -1;
+        if (a.minInterval > b.minInterval)
+            return 1;
+        return 0;
+    }
+    
+    private string GetPrettyCapability(OpenSeeWebcamCapability cap) {
+        float fps = 10000000f / (float)cap.minInterval;
+        string prettyCap = cap.minCX + "x" + cap.minCY + ", " + fps.ToString("0.##") + "fps (" + cap.format.ToString() + ")";
+        return prettyCap;
+    }
+    
+    public List<string> GetPrettyCapabilities() {
+        if (splitCaps != null && prettyCaps != null)
+            return prettyCaps;
+        
+        splitCaps = new List<Tuple<OpenSeeWebcamCapability, int>>();
+        
+        for (int i = 0; i < caps.Length; i++) {
+            if (caps[i].minCX == caps[i].maxCX && caps[i].minCY == caps[i].maxCY)
+                splitCaps.Add(new Tuple<OpenSeeWebcamCapability, int>(caps[i], i));
+            else {
+                OpenSeeWebcamCapability min = caps[i];
+                OpenSeeWebcamCapability max = caps[i];
+                min.maxCX = min.minCX;
+                min.maxCY = min.minCY;
+                max.minCX = max.maxCX;
+                max.minCY = max.maxCY;
+                splitCaps.Add(new Tuple<OpenSeeWebcamCapability, int>(min, i));
+                splitCaps.Add(new Tuple<OpenSeeWebcamCapability, int>(max, i));
+            }
+        }
+        
+        splitCaps.Sort((a, b) => CompareCaps(a.Item1, b.Item1));
+        
+        prettyCaps = new List<string>();
+        if (type == OpenSeeWebcamType.DirectShow)
+            prettyCaps.Add("Default settings");
+        foreach (var cap in splitCaps) {
+            prettyCaps.Add(GetPrettyCapability(cap.Item1));
+        }
+        
+        return prettyCaps;
+    }
+    
+    public OpenSeeWebcamCapability GetCapabilityByPrettyIndex(int index) {
+        int dshowOffset = 0;
+        if (type == OpenSeeWebcamType.DirectShow) {
+            if (index == 0) {
+                OpenSeeWebcamCapability cap = new OpenSeeWebcamCapability();
+                cap.id = -1;
+                cap.format = OpenSeeWebcamFormat.Any;
+                cap.minInterval = 10000000 / 60;
+                return cap;
+            }
+            dshowOffset = 1;
+        }
+        
+        if (splitCaps == null || index >= splitCaps.Count + dshowOffset || index < 0)
+            throw new InvalidOperationException("Invalid capability index.");
+        
+        return caps[splitCaps[index - dshowOffset].Item2];
+    }
+}
+
+[DefaultExecutionOrder(-50)]
+public class OpenSeeWebcamInfo : MonoBehaviour {
+    #region DllImport
+    // DirectShow
+    [DllImport("dshowcapture_x64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "create_capture")]
+	private static extern System.IntPtr create_capture_x64();
+
+    [DllImport("dshowcapture_x64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json_length")]
+	private static extern int get_json_length_x64(System.IntPtr cap);
+
+    [DllImport("dshowcapture_x64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json")]
+	private static extern void get_json_x64(System.IntPtr cap, [Out] StringBuilder namebuffer, int bufferlength);
+    
+    [DllImport("dshowcapture_x64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "destroy_capture")]
+	private static extern void destroy_capture_x64(System.IntPtr cap);
+
+    [DllImport("dshowcapture_x86", CallingConvention = CallingConvention.Cdecl, EntryPoint = "create_capture")]
+	private static extern System.IntPtr create_capture_x86();
+
+    [DllImport("dshowcapture_x86", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json_length")]
+	private static extern int get_json_length_x86(System.IntPtr cap);
+
+    [DllImport("dshowcapture_x86", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json")]
+	private static extern void get_json_x86(System.IntPtr cap, [Out] StringBuilder namebuffer, int bufferlength);
+
+    [DllImport("dshowcapture_x86", CallingConvention = CallingConvention.Cdecl, EntryPoint = "destroy_capture")]
+	private static extern void destroy_capture_x86(System.IntPtr cap);
+
+    // Blackmagic
+    [DllImport("libminibmcapture64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json_length")]
+	private static extern int bm_get_json_length_x64();
+
+    [DllImport("libminibmcapture64", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json")]
+	private static extern void bm_get_json_x64([Out] StringBuilder namebuffer, int bufferlength);
+
+    [DllImport("libminibmcapture32", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json_length")]
+	private static extern int bm_get_json_length_x86();
+
+    [DllImport("libminibmcapture32", CallingConvention = CallingConvention.Cdecl, EntryPoint = "get_json")]
+	private static extern void bm_get_json_x86([Out] StringBuilder namebuffer, int bufferlength);
+    #endregion
+    
+    [Serializable]
+    public class OpenSeeWebcamList {
+        public List<OpenSeeWebcam> list;
+    }
+
+    public List<OpenSeeWebcam> cameras;
+    public bool dumpJson = false;
+    
+    private static bool dumpJsonStatic = false;
+    
+    static private string ListCameraDetails_x64() {
+        System.IntPtr cap = create_capture_x64();
+		int length = get_json_length_x64(cap);
+        StringBuilder buffer = new StringBuilder(length);
+        get_json_x64(cap, buffer, length);
+        destroy_capture_x64(cap);
+		return buffer.ToString();
+    }
+
+    static private string ListCameraDetails_x86() {
+        System.IntPtr cap = create_capture_x86();
+		int length = get_json_length_x86(cap);
+        StringBuilder buffer = new StringBuilder(length);
+        get_json_x86(cap, buffer, length);
+        destroy_capture_x86(cap);
+		return buffer.ToString();
+    }
+
+    static private string ListBlackMagicDetails_x64() {
+		int length = bm_get_json_length_x64();
+        StringBuilder buffer = new StringBuilder(length);
+        bm_get_json_x64(buffer, length);
+		return buffer.ToString();
+    }
+
+    static private string ListBlackMagicDetails_x86() {
+		int length = bm_get_json_length_x86();
+        StringBuilder buffer = new StringBuilder(length);
+        bm_get_json_x86(buffer, length);
+		return buffer.ToString();
+    }
+
+    static public List<OpenSeeWebcam> ListCameraDetails(bool includeBlackMagic) {
+        string jsonData = null;
+        string bmJsonData = null;
+        if (Environment.Is64BitProcess) {
+            jsonData = ListCameraDetails_x64();
+            if (includeBlackMagic)
+                bmJsonData = ListBlackMagicDetails_x64();
+        } else {
+            jsonData = ListCameraDetails_x86();
+            if (includeBlackMagic)
+                bmJsonData = ListBlackMagicDetails_x86();
+        }
+        if (dumpJsonStatic) {
+            UnityEngine.Debug.Log("Camera JSON: " + jsonData);
+            if (includeBlackMagic)
+                UnityEngine.Debug.Log("Blackmagic JSON: " + bmJsonData);
+        }
+        List<OpenSeeWebcam> details = JsonUtility.FromJson<OpenSeeWebcamList>("{\"list\":" + jsonData + "}").list;
+        foreach (var cam in details)
+            cam.type = OpenSeeWebcamType.DirectShow;
+        if (includeBlackMagic) {
+            List<OpenSeeWebcam> bmDetails = JsonUtility.FromJson<OpenSeeWebcamList>("{\"list\":" + bmJsonData + "}").list;
+            foreach (var cam in bmDetails)
+                cam.type = OpenSeeWebcamType.Blackmagic;
+            details.AddRange(bmDetails);
+        }
+        return details;
+    }
+    
+    void Start() {
+        Initialize(false);
+    }
+    
+    public void Initialize(bool includeBlackMagic) {
+        dumpJsonStatic = dumpJson;
+        cameras = ListCameraDetails(includeBlackMagic);
+        foreach (var camera in cameras)
+            camera.GetPrettyCapabilities();
+    }
+}
+}

Examples/OpenSeeWebcamInfo.cs.meta

@@ -0,0 +1,11 @@
+fileFormatVersion: 2
+guid: 2e9598f1218283b418c7e72604be0335
+MonoImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  defaultReferences: []
+  executionOrder: -50
+  icon: {instanceID: 0}
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Licenses/ThunderSVM.txt

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright {yyyy} {name of copyright owner}
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

Licenses/libminibmcapture.txt

@@ -0,0 +1,19 @@
+zlib License
+
+libminibmcapture (C) 2020 noorus
+
+This software is provided 'as-is', without any express or implied
+warranty.  In no event will the authors be held liable for any damages
+arising from the use of this software.
+
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it
+freely, subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not
+   claim that you wrote the original software. If you use this software
+   in a product, an acknowledgment in the product documentation would be
+   appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be
+   misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.

Licenses/scikit-image.txt

@@ -0,0 +1,81 @@
+Copyright (C) 2019, the scikit-image team
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+ 1. Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+ 2. Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+ 3. Neither the name of skimage nor the names of its contributors may be
+    used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
+IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+
+skimage/_shared/version_requirements.py:_check_version
+
+    Copyright (c) 2013 The IPython Development Team
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice, this
+      list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+      this list of conditions and the following disclaimer in the documentation
+      and/or other materials provided with the distribution.
+
+    * Neither the name of the copyright holder nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+    AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+    IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+    DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+    FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+    DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+    SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+    CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+    OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+skimage/_shared/version_requirements.py:is_installed:
+
+    Original Copyright (C) 2009-2011 Pierre Raybaut
+
+    Permission is hereby granted, free of charge, to any person obtaining
+    a copy of this software and associated documentation files (the
+    "Software"), to deal in the Software without restriction, including
+    without limitation the rights to use, copy, modify, merge, publish,
+    distribute, sublicense, and/or sell copies of the Software, and to
+    permit persons to whom the Software is furnished to do so, subject to
+    the following conditions:
+
+    The above copyright notice and this permission notice shall be
+    included in all copies or substantial portions of the Software.
+
+    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+    NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+    LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+    OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+    WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -1,6 +1,8 @@
+![OSF.png](https://raw.githubusercontent.com/emilianavt/OpenSeeFace/master/Images/OSF.png)
+
 # Overview
 
-**Note**: This is a tracking library, **not** a stand-alone avatar puppeteering program. I'm also working on [such](https://www.vseeface.icu/) an application, which is currently in a semi-public beta, with releases being posted in the `#tech-dev` channel of [@Virtual_Deat](https://twitter.com/Virtual_Deat)'s Discord server.
+**Note**: This is a tracking library, **not** a stand-alone avatar puppeteering program. I'm also working on [VSeeFace](https://www.vseeface.icu/), which allows animating [VRM](https://vrm.dev/en/how_to_make_vrm/) and [VSFAvatar](https://www.youtube.com/watch?v=jhQ8DF87I5I) 3D models by using OpenSeeFace tracking. [VTube Studio](https://denchisoft.com/) uses OpenSeeFace for webcam based tracking to animate Live2D models. A renderer for the Godot engine can be found [here](https://github.com/virtual-puppet-project/vpuppr).
 
 This project implements a facial landmark detection model based on MobileNetV3.
 
@@ -8,15 +10,21 @@ As Pytorch 1.3 CPU inference speed on Windows is very low, the model was convert
 
 If anyone is curious, the name is a silly pun on the open seas and seeing faces. There's no deeper meaning.
 
-[Unity sample video](https://twitter.com/emiliana_vt/status/1210622149314203648) | [Sample video](https://www.youtube.com/watch?v=AOPHiAp9DBE) | [Sample video](https://www.youtube.com/watch?v=-cBSuHGdBWQ) | [Model comparison](https://www.youtube.com/watch?v=yMcJszUy7FA)
-
-These sample videos use older versions of the tracking models. Tracking accuracy of the current models is better.
-
 An up to date sample video can be found [here](https://www.youtube.com/watch?v=AaNap_ud_3I&vq=hd1080), showing the default tracking model's performance under different noise and light levels.
 
+# Tracking quality
+
+Since the landmarks used by OpenSeeFace are a bit different from those used by other approaches (they are close to iBUG 68, with two less points in the mouth corners and quasi-3D face contours instead of face contours that follow the visible outline) it is hard to numerically compare its accuracy to that of other approaches found commonly in scientific literature. The tracking performance is also more optimized for making landmarks that are useful for animating an avatar than for exactly fitting the face image. For example, as long as the eye landmarks show whether the eyes are opened or closed, even if their location is somewhat off, they can still be useful for this purpose.
+
+From general observation, OpenSeeFace performs well in adverse conditions (low light, high noise, low resolution) and keeps tracking faces through a very wide range of head poses with relatively high stability of landmark positions. Compared to MediaPipe, OpenSeeFace landmarks remain more stable in challenging conditions and it accurately represents a wider range of mouth poses. However, tracking of the eye region can be less accurate.
+
+I ran OpenSeeFace on a sample clip from the video presentation for [3D Face Reconstruction with Dense Landmarks](https://microsoft.github.io/DenseLandmarks/) by Wood et al. to compare it to MediaPipe and their approach. You can watch the result [here](https://www.vseeface.icu/assets/media/OSFMediaPipe3DFR.mp4).
+
 # Usage
 
-The face tracking itself is done by the `facetracker.py` Python 3.7 script. It is a commandline program, so you should start it manually from cmd or write a batch file to start it.
+A sample Unity project for VRM based avatar animation can be found [here](https://github.com/emilianavt/OpenSeeFaceSample).
+
+The face tracking itself is done by the `facetracker.py` Python 3.7 script. It is a commandline program, so you should start it manually from cmd or write a batch file to start it. If you downloaded a release and are on Windows, you can run the `facetracker.exe` inside the `Binary` folder without having Python installed. You can also use the `run.bat` inside the `Binary` folder for a basic demonstration of the tracker.
 
 The script will perform the tracking on webcam input or video file and send the tracking data over UDP. This design also allows tracking to be done on a separate PC from the one who uses the tracking information. This can be useful to enhance performance and to avoid accidentially revealing camera footage.
 
@@ -30,6 +38,8 @@ A simple demonstration can be achieved by creating a new scene in Unity, adding
 
     python facetracker.py --visualize 3 --pnp-points 1 --max-threads 4 -c video.mp4
 
+__Note__: If dependencies were installed using [poetry](https://python-poetry.org/), the commands have to be executed from a `poetry shell` or have to be prefixed with `poetry run`.
+
 This way the tracking script will output its own tracking visualization while also demonstrating the transmission of tracking data to Unity.
 
 The included `OpenSeeLauncher` component allows starting the face tracker program from Unity. It is designed to work with the pyinstaller created executable distributed in the binary release bundles. It provides three public API functions:
@@ -83,6 +93,7 @@ To save both the trained model and the captured training data, type in a filenam
 
 Four pretrained face landmark models are included. Using the `--model` switch, it is possible to select them for tracking. The given fps values are for running the model on a single face video on a single CPU core. Lowering the frame rate would reduce CPU usage by a corresponding degree.
 
+* Model **-1**: This model is for running on toasters, so it's a very very fast and very low accuracy model. (213fps without gaze tracking)
 * Model **0**: This is a very fast, low accuracy model. (68fps)
 * Model **1**: This is a slightly slower model with better accuracy. (59fps)
 * Model **2**: This is a slower model with good accuracy. (50fps)
@@ -90,7 +101,7 @@ Four pretrained face landmark models are included. Using the `--model` switch, i
 
 FPS measurements are from running on one core of my CPU.
 
-Pytorch weights for use with `model.py` can be found [here](https://mega.nz/file/vvYXlYQT#h7FpEg4tmOCJNxjpsDEw0JomJIkVGKwrt4OUV0RNDDU).
+Pytorch weights for use with `model.py` can be found [here](https://mega.nz/file/vvYXlYQT#h7FpEg4tmOCJNxjpsDEw0JomJIkVGKwrt4OUV0RNDDU). Some unoptimized ONNX models can be found [here](https://github.com/emilianavt/OpenSeeFace/issues/48).
 
 # Results
 
@@ -118,9 +129,8 @@ When distributing it, you should also distribute the `Licenses` folder along wit
 
 The release builds contain a custom build of ONNX Runtime without telemetry.
 
-# Dependencies
+# Dependencies (Python 3.6 - 3.9)
 
-* Python 3.7
 * ONNX Runtime
 * OpenCV
 * Pillow
@@ -130,6 +140,22 @@ The required libraries can be installed using pip:
 
      pip install onnxruntime opencv-python pillow numpy
 
+Alternatively poetry can be used to 
+install all dependencies for this project in a separate virtual env:
+
+     poetry install
+
+# Dependencies
+
+* onnxruntime
+* OpenCV
+* Pillow
+* Numpy
+
+The required libraries can be installed using pip:
+ 
+	pip install onnxruntime opencv-python pillow numpy
+
 # References
 
 ## Training dataset
@@ -155,6 +181,8 @@ Additional training has been done on the WFLW dataset after reducing it to 66 po
 
 For the training the gaze and blink detection model, the [MPIIGaze](https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/gaze-based-human-computer-interaction/appearance-based-gaze-estimation-in-the-wild/) dataset was used. Additionally, around 125000 synthetic eyes generated with [UnityEyes](https://www.cl.cam.ac.uk/research/rainbow/projects/unityeyes/) were used during training.
 
+It should be noted that additional custom data was also used during the training process and that the reference landmarks from the original datasets have been modified in certain ways to address various issues. It is likely not possible to reproduce these models with just the original LS3D-W and WFLW datasets, however the additional data is not redistributable.
+
 The heatmap regression based face detection model was trained on random 224x224 crops from the WIDER FACE dataset.
 
 	@inproceedings{yang2016wider,

Source/ThunderSVMWrapper.cpp

@@ -0,0 +1,222 @@
+// ThunderSVMWrapper.cpp : Defines the exported functions for the DLL.
+//
+
+#include "pch.h"
+#include "framework.h"
+#include "ThunderSVMWrapper.h"
+#include "thundersvm/thundersvm.h"
+#include "thundersvm/model/svmmodel.h"
+#include "thundersvm/model/svc.h"
+#include <cmath>
+
+using namespace std;
+
+extern "C" void __cdecl print_log(const char *str) {
+    return;
+}
+
+struct problem {
+    int r;
+    int c;
+    DataSet::node2d x;
+    vector<double> y;
+    DataSet dataset;
+};
+
+struct SVMWrapper {
+    SVC *model;
+    int cols;
+    int classes;
+    double *means;
+    double *sdevs;
+    SvmParam param;
+    struct problem problem;
+};
+
+typedef struct svm_node svmnode;
+
+void make_svmnode_array(DataSet::node2d &dataset, float features[], int rows, int cols, double *means, double *sdevs, int rescale) {
+    double row_factor = 1.0 / (double)rows;
+    if (rescale) {
+        for (int c = 0; c < cols; c++) {
+            means[c] = 0;
+            sdevs[c] = 0;
+        }
+        for (int r = 0; r < rows; r++) {
+            for (int c = 0; c < cols; c++) {
+                means[c] += (double)features[r * cols + c];
+            }
+        }
+        for (int c = 0; c < cols; c++) {
+            means[c] *= row_factor;
+        }
+        for (int r = 0; r < rows; r++) {
+            for (int c = 0; c < cols; c++) {
+                double sq_diff = ((double)features[r * cols + c] - means[c]);
+                sq_diff *= sq_diff;
+                sdevs[c] += sq_diff;
+            }
+        }
+        for (int c = 0; c < cols; c++) {
+            sdevs[c] *= row_factor;
+            sdevs[c] = sqrt(sdevs[c]);
+            sdevs[c] = 1.0 / sdevs[c];
+        }
+    }
+
+    for (int r = 0; r < rows; r++) {
+        dataset[r] = vector<DataSet::node>(cols);
+        int idx = 0;
+        for (int c = 0; c < cols; c++) {
+            dataset[r][c].index = c + 1;
+            dataset[r][c].value = (float)((features[r * cols + c] - means[c]) * sdevs[c]);
+        }
+    }
+}
+
+void make_problem(struct problem &problem, float features[], float labels[], int rows, int cols, double *means, double *sdevs) {
+    problem.r = rows;
+    problem.c = cols;
+    problem.y = vector<double>(rows);
+    for (int i = 0; i < rows; i++)
+        problem.y[i] = (double)labels[i];
+    problem.x = DataSet::node2d(rows);
+    make_svmnode_array(problem.x, features, rows, cols, means, sdevs, 1);
+    problem.dataset = DataSet(problem.x, cols, problem.y);
+}
+
+void destroy_problem(struct problem &problem) {
+    problem.y.clear();
+    problem.x.clear();
+}
+
+extern "C" {
+    __declspec(dllexport) void *__cdecl trainModel(float features[], float labels[], float weights[], int rows, int cols, int classes, int probability, float C) {
+        if (!(rows > 0 && cols > 0))
+            return NULL;
+
+        //svm_set_print_string_function(print_log);
+
+        struct SVMWrapper *sw = new SVMWrapper();
+        sw->cols = cols;
+        sw->classes = classes;
+        sw->means = new double[cols];
+        sw->sdevs = new double[cols];
+
+        sw->param.svm_type = SvmParam::C_SVC;
+        sw->param.kernel_type = SvmParam::RBF;
+        sw->param.degree = 3;
+        sw->param.gamma = 1.0 / (double)cols;
+        sw->param.coef0 = 0;
+        sw->param.nu = 0.5;
+        sw->param.C = C;
+        sw->param.epsilon = 0.001;
+        sw->param.p = 0.1;
+        sw->param.probability = probability;
+        sw->param.nr_weight = 0;
+        sw->param.weight_label = NULL;
+        sw->param.weight = NULL;
+
+        if (weights != NULL) {
+            sw->param.nr_weight = classes;
+            sw->param.weight_label = new int[classes];
+            sw->param.weight = new double[classes];
+            for (int i = 0; i < classes; i++) {
+                sw->param.weight_label[i] = i;
+                sw->param.weight[i] = (double)weights[i];
+            }
+        }
+
+        make_problem(sw->problem, features, labels, rows, cols, sw->means, sw->sdevs);
+        sw->model = new SVC();//svm_train(sw->problem, &sw->param);
+        sw->model->train(sw->problem.dataset, sw->param);
+
+        if (weights != NULL) {
+            delete sw->param.weight_label;
+            delete sw->param.weight;
+        }
+
+        return sw;
+    }
+
+    __declspec(dllexport) void __cdecl predict(void *ptr, float features[], float predictions[], double probabilities[], int rows) {
+        struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
+        int cols = sw->cols;
+
+        DataSet::node2d instances = DataSet::node2d(rows);
+        make_svmnode_array(instances, features, rows, sw->cols, sw->means, sw->sdevs, 0);
+        vector<double> predicted = sw->model->predict(instances, -1);
+        instances.clear();
+        for (int r = 0; r < rows; r++)
+            predictions[r] = (float)predicted[r];
+        if (probabilities != NULL && sw->param.probability) {
+            vector<float> probs = sw->model->get_prob_predict();
+            for (int r = 0; r < rows; r++)
+                for (int c = 0; c < sw->classes; c++)
+                    probabilities[c + r * sw->classes] = probs[c + r * sw->classes];
+        }
+        predicted.clear();
+
+        return;
+    }
+
+    __declspec(dllexport) void *__cdecl loadModel(char *filename, int cols, int classes, double means[], double sdevs[]) {
+        struct SVMWrapper *sw = new SVMWrapper();
+        sw->cols = cols;
+        sw->model = new SVC();
+        sw->model->load_from_file(filename);
+        sw->classes = classes;
+        sw->means = new double[cols];
+        sw->sdevs = new double[cols];
+        for (int c = 0; c < cols; c++) {
+            sw->means[c] = means[c];
+            sw->sdevs[c] = sdevs[c];
+        }
+        return sw;
+    }
+
+    __declspec(dllexport) void *__cdecl loadModelString(char *modelString, int cols, int classes, double means[], double sdevs[]) {
+        struct SVMWrapper *sw = new SVMWrapper();
+        sw->cols = cols;
+        sw->model = new SVC();
+        sw->model->load_from_string(modelString);
+        sw->classes = classes;
+        sw->means = new double[cols];
+        sw->sdevs = new double[cols];
+        for (int c = 0; c < cols; c++) {
+            sw->means[c] = means[c];
+            sw->sdevs[c] = sdevs[c];
+        }
+        return sw;
+    }
+
+    __declspec(dllexport) void __cdecl getScales(void *ptr, double means[], double sdevs[]) {
+        struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
+        int cols = sw->cols;
+        for (int c = 0; c < cols; c++) {
+            means[c] = sw->means[c];
+            sdevs[c] = sw->sdevs[c];
+        }
+    }
+
+    __declspec(dllexport) const char *__cdecl saveModelString(void *ptr) {
+        struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
+        string model = sw->model->save_to_string();
+        return _strdup(model.c_str());
+    }
+
+    __declspec(dllexport) int __cdecl saveModel(void *ptr, char *filename) {
+        struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
+        sw->model->save_to_file(string(filename));
+        return 1;
+    }
+
+    __declspec(dllexport) void __cdecl destroyModel(void *ptr) {
+        struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
+        delete sw->means;
+        delete sw->sdevs;
+        delete sw->model;
+        destroy_problem(sw->problem);
+        delete sw;
+    }
+}

Unity/OpenSee.cs

@@ -24,6 +24,8 @@ public class OpenSee : MonoBehaviour {
     [Tooltip("This contains the actual tracking data")]
     public OpenSeeData[] trackingData = null;
     
+    public bool listening { get; private set; } = false;
+    
     [HideInInspector]
     public float maxFit3DError = 100f;
 
@@ -39,9 +41,9 @@ public class OpenSee : MonoBehaviour {
         public float rightEyeOpen;
         [Tooltip("This field tells you how likely it is that the left eye is open.")]
         public float leftEyeOpen;
-        [Tooltip("This field tells you how likely it is that the right eye is open.")]
+        [Tooltip("This field contains the rotation of the right eyeball.")]
         public Quaternion rightGaze;
-        [Tooltip("This field tells you how likely it is that the left eye is open.")]
+        [Tooltip("This field contains the rotation of the left eyeball.")]
         public Quaternion leftGaze;
         [Tooltip("This field tells you if 3D points have been successfully estimated from the 2D points. If this is false, do not rely on pose or 3D data.")]
         public bool got3DPoints;
@@ -211,6 +213,7 @@ public class OpenSee : MonoBehaviour {
 
     void performReception() {
         EndPoint senderRemote = new IPEndPoint(IPAddress.Any, 0);
+        listening = true;
         while (!stopReception) {
             try {
                 int receivedBytes = socket.ReceiveFrom(buffer, SocketFlags.None, ref senderRemote);

Unity/OpenSeeExpression.cs

@@ -13,6 +13,18 @@ public class OpenSeeExpression : MonoBehaviour
     [Header("Settings")]
     [Tooltip("This is the source of face tracking data. If it is not set, any OpenSee component on the current object is used.")]
     public OpenSee openSee;
+    [Tooltip("When enabled ThunderSVM is used instead of LibSVM, which is usually faster, but may cause issues on some systems.")]
+    public bool useThunderSVM = false;
+    [Header("Simple mode")]
+    [Tooltip("When enabled, some simple heuristics will be used to detect expressions, rather than an SVM model. Outputs one of: neutral, fun, angry, surprise")]
+    public bool simpleMode = false;
+    [Tooltip("This smoothing factor is applied to the features used for simple expression detection.")]
+    [Range(0f, 1f)]
+    public float simpleSmoothing = 0.6f;
+    [Tooltip("This smoothing factor is applied to the features used for simple expression detection.")]
+    [Range(0f, 5f)]
+    public float simpleSensitivity = 1f;
+    [Header("Calibration")]
     [Tooltip("This specifies which expression calibration data should be collected for.")]
     public string calibrationExpression = "neutral";
     [Tooltip("This specifies the id of the face for which data should be collected and predictions should be made. Face ids depend only on the order of first detection and locations of the faces.")]
@@ -50,7 +62,7 @@ public class OpenSeeExpression : MonoBehaviour
     [Tooltip("These settings can be used to train an expression detection model on a subset of points.")]
     public PointSelection pointSelection;
     [Header("Information")]
-    [Tooltip("This is the of expressions for which calibration data was collected. The maximum number is 10.")]
+    [Tooltip("This is the number of expressions for which calibration data was collected. The maximum number is 10.")]
     public int expressionNumber = 0;
     [Tooltip("This is the percentage of necessary training data collected for the current expression.")]
     public float percentRecorded = 0.0f;
@@ -72,6 +84,44 @@ public class OpenSeeExpression : MonoBehaviour
     [Tooltip("This shows any accuracy warnings that resulted from the last training run.")]
     public string[] warnings = null;
     
+    private float lastMouthCorner = 0f;
+    private float lastEyebrows = 0f;
+    private bool hadFun = false;
+    private bool hadAngry = false;
+    private bool hadSurprised = false;
+    float AdjustThreshold(bool active) {
+        if (active)
+            return 0.8f;
+        else
+            return 1f;
+    }
+    void ThresholdDetection() {
+        lastMouthCorner = lastMouthCorner * simpleSmoothing + (openSeeData.features.MouthCornerUpDownLeft + openSeeData.features.MouthCornerUpDownRight) * 0.5f * (1f - simpleSmoothing);
+        lastEyebrows = lastEyebrows * simpleSmoothing + (openSeeData.features.EyebrowUpDownLeft + openSeeData.features.EyebrowUpDownRight) * 0.5f * (1f - simpleSmoothing);
+        if (lastMouthCorner * simpleSensitivity < -0.2f * AdjustThreshold(hadFun)) {
+            expression = "fun";
+            hadFun = true;
+            hadSurprised = false;
+            hadAngry = false;
+        } else if (lastEyebrows * simpleSensitivity > 0.2f * AdjustThreshold(hadSurprised)) {
+            expression = "surprise";
+            hadFun = false;
+            hadSurprised = true;
+            hadAngry = false;
+        } else if (lastEyebrows * simpleSensitivity < -0.25f * AdjustThreshold(hadAngry) && lastMouthCorner * simpleSensitivity > -0.3f * (2f - AdjustThreshold(hadAngry))) {
+            expression = "angry";
+            hadFun = false;
+            hadSurprised = false;
+            hadAngry = true;
+        } else {
+            expression = "neutral";
+            hadFun = false;
+            hadSurprised = false;
+            hadAngry = false;
+        }
+        expressionTime = Time.time;
+    }
+
     [Serializable]
     public class PointSelection {
         [Tooltip("When enabled, the points of the face contour will be used to determine the expression.")]
@@ -104,6 +154,7 @@ public class OpenSeeExpression : MonoBehaviour
         private string[] classLabels = null;
         private int[] indices = null;
         private PointSelection pointSelection;
+        private bool thunderSVM = false;
         //private bool newModel = false;
 
         static public void LoadSerialized(byte[] modelBytes, out Dictionary<string, List<float[]>> expressions, out SVMModel model, out string[] classLabels, out int[] indices, ref PointSelection pointSelection) {
@@ -115,6 +166,9 @@ public class OpenSeeExpression : MonoBehaviour
                 oser = formatter.Deserialize(gzipStream) as OpenSeeExpressionRepresentation;
             }
             expressions = oser.expressions;
+            /*if (oser.thunderSVM)
+                model = new ThunderSVMModel(oser.modelBytes);
+            else*/
             model = new SVMModel(oser.modelBytes);
             classLabels = oser.classLabels;
             indices = oser.indices;
@@ -137,6 +191,7 @@ public class OpenSeeExpression : MonoBehaviour
             oser.classLabels = classLabels;
             oser.indices = indices;
             oser.pointSelection = pointSelection;
+            oser.thunderSVM = model is ThunderSVMModel;
             //oser.newModel = true;
 
             IFormatter formatter = new BinaryFormatter();
@@ -309,7 +364,10 @@ public class OpenSeeExpression : MonoBehaviour
         }
         ResetInfo();
         expressions = new Dictionary<string, List<float[]>>();
-        model = new SVMModel();
+        if (useThunderSVM)
+            model = new ThunderSVMModel();
+        else
+            model = new SVMModel();
         rnd = new System.Random();
     }
 
@@ -363,7 +421,10 @@ public class OpenSeeExpression : MonoBehaviour
             return false;
         }
         modelReady = false;
-        //model = new SVMModel();
+        if (model is SVMModel && useThunderSVM)
+            model = new ThunderSVMModel();
+        else if (model is ThunderSVMModel && !useThunderSVM)
+            model = new SVMModel();
         SelectPoints();
         List<string> keys = new List<string>();
         List<string> accuracyWarnings = new List<string>();
@@ -377,13 +438,13 @@ public class OpenSeeExpression : MonoBehaviour
                 samples += list.Count;
             } else {
                 if (list != null && list.Count > 20) {
-                    Debug.Log("[Training warning] Expression " + key + " has little data and might be inaccurate.");
+                    Debug.Log("[Training warning] Expression " + key + " has little data and might be inaccurate. (" + list.Count + ")");
-                    accuracyWarnings.Add("Expression " + key + " has little data and might be inaccurate.");
+                    accuracyWarnings.Add("Expression " + key + " has little data and might be inaccurate. (" + list.Count + ")");
                     samples += list.Count;
                     keys.Add(key);
                 } else {
-                    Debug.Log("[Training warning] Skipping expression " + key + " due to lack of collected data.");
+                    Debug.Log("[Training warning] Skipping expression " + key + " due to lack of collected data. (" + list.Count + ")");
-                    accuracyWarnings.Add("Skipping expression " + key + " due to lack of collected data.");
+                    accuracyWarnings.Add("Skipping expression " + key + " due to lack of collected data. (" + list.Count + ")");
                     continue;
                 }
             }
@@ -401,11 +462,16 @@ public class OpenSeeExpression : MonoBehaviour
         if (weightMap != null) {
             weights = new float[classLabels.Length];
             for (int i = 0; i < classLabels.Length; i++) {
-                weights[i] = weightMap[classLabels[i]];
+                if (weightMap.ContainsKey(classLabels[i]))
+                    weights[i] = weightMap[classLabels[i]];
+                else
+                    weights[i] = 1f;
                 Debug.Log("[Training info] Adding weight " + weights[i] + " for " + classLabels[i] + ".");
             }
         }
         
+        Debug.Log("[Training info] Preparing trainig data.");
+        
         int train_split = maxSamples * 3 / 4;
         int test_split = maxSamples - train_split;
         int rows_train = classes * train_split;
@@ -455,6 +521,7 @@ public class OpenSeeExpression : MonoBehaviour
                 i_test++;
             }
         }
+        Debug.Log("[Training info] Preparation complete.");
         int probability = 0;
         if (enableProbabilityTraining)
             probability = 1;
@@ -474,17 +541,24 @@ public class OpenSeeExpression : MonoBehaviour
         }
         warnings = accuracyWarnings.ToArray();
         modelReady = true;
+        Debug.Log("[Training info] Trained model.");
         return true;
     }
 
     public bool PredictExpression() {
-        if (openSee == null || !(modelReady && model.Ready())) {
+        if (openSee == null)
+            return false;
+        if (!simpleMode && !(modelReady && model.Ready())) {
             ResetInfo();
             return false;
         }
         openSeeData = openSee.GetOpenSeeData(faceId);
         if (openSeeData == null || openSeeData.time <= lastCapture)
             return false;
+        if (simpleMode) {
+            ThresholdDetection();
+            return true;
+        }
         float[] faceData = GetData(openSeeData);
         float[] predictionData = new float[cols];
         for (int i = 0; i < cols; i++)
@@ -632,7 +706,7 @@ public class OpenSeeExpression : MonoBehaviour
             SaveToFile();
         if (!predict)
             expression = "";
-        else if (modelReady)
+        else if (simpleMode || modelReady)
             PredictExpression();
         else
             predict = false;

Unity/OpenSeeIKTarget.cs

@@ -2,17 +2,24 @@
 
 namespace OpenSee {
 
+[DefaultExecutionOrder(-1)]
 public class OpenSeeIKTarget : MonoBehaviour
 {
     [Header("Settings")]
     [Tooltip("This is the source of face tracking data.")]
     public OpenSee openSee;
+    [Tooltip("When this is set to another game object, its kinematic interpolation component will get updated from values of this.")]
+    public OpenSeeKinematicInterpolation kinematicInterpolation;
+    [Tooltip("When set, the this object will follow the specified transform instead and automatically recalibrate based on it.")]
+    public Transform externalTarget = null;
     [Tooltip("This can be used to skip the first few tracking frames.")]
     public int skipFirst = 5;
     [Tooltip("When set to true, the next rotation and translation values will be set as zero rotation and zero translation.")]
     public bool calibrate = true;
     [Tooltip("This sets which face id to look for in the OpenSee data.")]
     public int faceId = 0;
+    [Tooltip("When enabled, the IK target will not move up and down.")]
+    public bool preventVerticalMotion = false;
     [Tooltip("When this is enabled, rotations and movement are mirrored.")]
     public bool mirrorMotion = false;
     [Tooltip("Often, the translation vector's scale is too high. Setting this somewhere between 0.05 to 0.3 seems stabilize things, but it also reduces the range of motion.")]
@@ -80,17 +87,46 @@ public class OpenSeeIKTarget : MonoBehaviour
     private bool skippedPosition = false;
     private float skipStartTimePosition = -1f;
     
+    private Transform lastExternalTarget = null;
+    
     public void Calibrate() {
         calibrate = true;
     }
     
+    public static Vector3 PrettyEuler(Vector3 last, Vector3 current) {
+        // Ew, but how would you do componentwise filtering/interpolation with quaternions?
+        if (Mathf.Abs(current.x - last.x) > Mathf.Abs((current.x - 360f) - last.x))
+            current.x -= 360f;
+        else if (Mathf.Abs(current.x - last.x) > Mathf.Abs((current.x + 360f) - last.x))
+            current.x += 360f;
+        if (Mathf.Abs(current.y - last.y) > Mathf.Abs((current.y - 360f) - last.y))
+            current.y -= 360f;
+        else if (Mathf.Abs(current.y - last.y) > Mathf.Abs((current.y + 360f) - last.y))
+            current.x += 360f;
+        if (Mathf.Abs(current.z - last.z) > Mathf.Abs((current.z - 360f) - last.z))
+            current.z -= 360f;
+        else if (Mathf.Abs(current.z - last.z) > Mathf.Abs((current.z + 360f) - last.z))
+            current.z += 360f;
+        return current;
+    }
+    
+    Vector3 FilterPos(Vector3 a, Vector3 b) {
+        return Vector3.Lerp(a, b, 1f - smooth);
+    }
+    
+    Quaternion FilterRot(Quaternion a, Quaternion b) {
+        return Quaternion.Lerp(a, b, 1f - smooth);
+    }
+
     void Interpolate() {
         if (!interpolate || interpolateState < 2)
             return;
-        float t = Mathf.Clamp((float)interpolationCount / avgInterps, 0f, 0.985f);
+        float t = Mathf.Clamp((float)interpolationCount / avgInterps, 0f, 1.5f);
-        transform.localPosition = Vector3.Lerp(lastT, currentT, t);
+        transform.localPosition = Vector3.SlerpUnclamped(lastT, currentT, t);
-        transform.localRotation = Quaternion.Lerp(lastR, currentR, t);
+        transform.localRotation = Quaternion.SlerpUnclamped(lastR, currentR, t);
         interpolationCount++;
+        //transform.localPosition = new Vector3(0.3f * ((Time.time/3f) % 3f) - 0.5f, 0.3f * Mathf.Sin(Mathf.Rad2Deg * ((Time.time/5f) % 3f)), 0f);
+        //transform.localRotation = Quaternion.identity * Quaternion.AngleAxis(20f * Mathf.Sin(Mathf.Rad2Deg * ((Time.time/5f) % 3f)) - 10f, Vector3.up);
     }
 
     Quaternion MirrorQuaternion(Quaternion q) {
@@ -103,23 +139,33 @@ public class OpenSeeIKTarget : MonoBehaviour
 
     void RunUpdate() {
         var openSeeData = openSee.GetOpenSeeData(faceId);
-        if (openSeeData == null || openSeeData.fit3DError > openSee.maxFit3DError)
+        Quaternion convertedQuaternion;
-            return;
+        Vector3 t;
-        if (openSeeData.time > updated) {
+        Vector3 convertedTranslation;
-            updated = openSeeData.time;
-            trackingFrames++;
-            if (trackingFrames < skipFirst)
-                return;
-        } else {
-            Interpolate();
-            return;
-        }
         
-        Quaternion convertedQuaternion = new Quaternion(-openSeeData.rawQuaternion.y, -openSeeData.rawQuaternion.x, openSeeData.rawQuaternion.z, openSeeData.rawQuaternion.w);
+        if (externalTarget == null) {
-        Vector3 t = openSeeData.translation;
+            if (openSeeData == null || openSeeData.fit3DError > openSee.maxFit3DError)
-        t.x = -t.x;
+                return;
-        t.z = -t.z;
+            if (openSeeData.time > updated) {
-        Vector3 convertedTranslation = new Vector3(t.x, t.y, t.z);
+                updated = openSeeData.time;
+                trackingFrames++;
+                if (trackingFrames < skipFirst)
+                    return;
+            } else {
+                Interpolate();
+                return;
+            }
+            
+            convertedQuaternion = new Quaternion(-openSeeData.rawQuaternion.y, -openSeeData.rawQuaternion.x, openSeeData.rawQuaternion.z, openSeeData.rawQuaternion.w);
+            t = openSeeData.translation;
+            t.x = -t.x;
+            t.z = -t.z;
+            convertedTranslation = new Vector3(t.x, t.y, t.z);
+        } else {
+            convertedQuaternion = externalTarget.localRotation;
+            convertedTranslation = externalTarget.localPosition;
+            t = convertedTranslation;
+        }
         
         // Check for angular outliers
         if (gotAccepted) {
@@ -185,6 +231,9 @@ public class OpenSeeIKTarget : MonoBehaviour
             return;
         }
         
+        calibrate = calibrate || (externalTarget != lastExternalTarget);
+        lastExternalTarget = externalTarget;
+
         if (calibrate) {
             dR = convertedQuaternion;
             dR = Quaternion.Inverse(dR);
@@ -216,11 +265,11 @@ public class OpenSeeIKTarget : MonoBehaviour
         interpolationCount = 0;
         averageInterpolations = avgInterps;
         
-        lastT = currentT;
+        lastT = transform.localPosition;
-        lastR = currentR;
+        lastR = transform.localRotation;
         if (interpolateState > 0) {
-            currentT = Vector3.Lerp(currentT, (t - dT) * translationScale, 1f - smooth);
+            currentT = FilterPos(transform.localPosition, (t - dT) * translationScale);
-            currentR = Quaternion.Lerp(transform.localRotation, convertedQuaternion * dR, 1f - smooth);
+            currentR = FilterRot(transform.localRotation, convertedQuaternion * dR);
         } else {
             currentT = (t - dT) * translationScale;
             currentR = convertedQuaternion * dR;
@@ -231,8 +280,22 @@ public class OpenSeeIKTarget : MonoBehaviour
         if (interpolate) {
             Interpolate();
         } else {
-            transform.localPosition = Vector3.Lerp(transform.localPosition, (t - dT) * translationScale, 1f - smooth);
+            transform.localPosition = FilterPos(transform.localPosition, (t - dT) * translationScale);
-            transform.localRotation = Quaternion.Lerp(transform.localRotation, convertedQuaternion * dR, 1f - smooth);
+            transform.localRotation = FilterRot(transform.localRotation, convertedQuaternion * dR);
+        }
+
+       if (preventVerticalMotion) {
+           currentT.y = 0f;
+            Vector3 pos = transform.localPosition;
+            pos.y = 0f;
+            transform.localPosition = pos;
+        }
+
+        if (kinematicInterpolation != null && kinematicInterpolation.gameObject != gameObject) {
+            if (interpolate)
+                kinematicInterpolation.UpdateKI(updated, currentT, currentR);
+            else
+                kinematicInterpolation.UpdateKI(updated, transform.localPosition, transform.localRotation);
         }
         
         if (driftBack) {
@@ -248,6 +311,10 @@ public class OpenSeeIKTarget : MonoBehaviour
        }
     }
     
+    public void SetPreventVerticalMotion(bool v) {
+        preventVerticalMotion = v;
+    }
+
     void FixedUpdate()
     {
         if (fixedUpdate)

Unity/OpenSeeIKTarget.cs.meta

@@ -4,7 +4,7 @@ MonoImporter:
   externalObjects: {}
   serializedVersion: 2
   defaultReferences: []
-  executionOrder: 0
+  executionOrder: -5
   icon: {instanceID: 0}
   userData: 
   assetBundleName: 

Unity/OpenSeeInterpolation.cs

@@ -0,0 +1,232 @@
+using System.Collections;
+using System.Collections.Generic;
+using UnityEngine;
+
+namespace OpenSee {
+    public class InterpolatedMap<T, V, U> where V : InterpolatedValue<U>, new() {
+        private Dictionary<T, V> map = new Dictionary<T, V>();
+        private float clamp = 1f;
+        private float smoothing = 0f;
+        private bool gotDefault = false;
+        private U defaultValue;
+
+        public void Clear() {
+            map.Clear();
+        }
+        
+        public void SetSmoothing(float smoothing) {
+            this.smoothing = smoothing;
+            foreach (var entry in map) {
+                if (entry.Value != null)
+                    entry.Value.SetSmoothing(smoothing);
+            }
+        }
+
+        public void SetClamp(float clamp) {
+            this.clamp = clamp;
+            foreach (var entry in map) {
+                if (entry.Value != null)
+                    entry.Value.SetClamp(clamp);
+            }
+        }
+        
+        void Prepare(T key) {
+            if (!map.ContainsKey(key)) {
+                map.Add(key, new V());
+                map[key].SetSmoothing(smoothing);
+                map[key].SetClamp(clamp);
+            }
+        }
+        
+        public void SetValue(T key, U v) {
+            Prepare(key);
+            map[key].SetValue(v);
+        }
+        
+        public void Store(T key, U v) {
+            Prepare(key);
+            map[key].UpdateValue(v);
+        }
+
+        public void Store(T key, U v, double nowT) {
+            if (!map.ContainsKey(key)) {
+                map.Add(key, new V());
+                map[key].SetSmoothing(smoothing);
+                map[key].SetClamp(clamp);
+            }
+            map[key].UpdateValue(v, nowT);
+        }
+        
+        public U GetDefault() {
+            if (gotDefault)
+                return defaultValue;
+            V tmp = new V();
+            defaultValue = tmp.GetDefault();
+            return defaultValue;
+        }
+        
+        public bool Check(T key) {
+            return map.ContainsKey(key);
+        }
+        
+        public U Get(T key) {
+            if (!map.ContainsKey(key))
+                return GetDefault();
+            return map[key].Interpolate();
+        }
+    }
+
+    abstract public class InterpolatedValue<V> {
+        private float interpolateT = 0f;
+        private float updateT = 0f;
+        private double lastT = 0f;
+        private double currentT = 0f;
+        private float clamp = 1f;
+        private float smoothing = 0f;
+        private int gotData = 0;
+
+        private bool haveDefault = false;
+        private V defaultValue;
+
+        private bool valueSet = false;
+        private V dataValue;
+        private V dataValueLast;
+
+        abstract public V ValueLerpUnclamped(V a, V b, float t);
+        abstract public V GetBuiltinDefault();
+        
+        public void SetDefault(V defaultValue) {
+            this.defaultValue = defaultValue;
+            haveDefault = true;
+        }
+        
+        public void ResetDefault() {
+            haveDefault = false;
+        }
+        
+        public V GetDefault() {
+            if (haveDefault)
+                return defaultValue;
+            return GetBuiltinDefault();
+        }
+        
+        public void SetSmoothing(float smoothing) {
+            this.smoothing = smoothing;
+        }
+
+        public void SetClamp(float clamp) {
+            this.clamp = clamp;
+        }
+        
+        public void SetValue(V v) {
+            dataValueLast = v;
+            dataValue = v;
+            valueSet = true;
+        }
+        
+        public InterpolatedValue(float smoothing = 0f, float clamp = 1f) {
+            this.smoothing = smoothing;
+            this.clamp = clamp;
+        }
+        
+        public void UpdateValue(V v) {
+            UpdateValue(v, Time.realtimeSinceStartup);
+        }
+        
+        public void UpdateValue(V v, double nowT) {
+            if (nowT < 0f)
+                return;
+            lock (this) {
+                if (nowT < currentT || gotData == 0) {
+                    // Reinitialize
+                    gotData = 1;
+                    updateT = Time.time;
+                    lastT = 0f;
+                    currentT = nowT;
+                    dataValueLast = GetDefault();
+                    dataValue = ValueLerpUnclamped(v, dataValueLast, smoothing);
+                } else if (nowT == currentT) {
+                    // Change value
+                    dataValue = ValueLerpUnclamped(v, dataValueLast, smoothing);
+                } else {
+                    // Regular operation
+                    updateT = Time.time;
+                    lastT = currentT;
+                    currentT = nowT;
+                    if (gotData < 2)
+                        gotData++;
+                    dataValueLast = dataValue;
+                    dataValue = ValueLerpUnclamped(v, dataValueLast, smoothing);
+                }
+            }
+        }
+        
+        public V Interpolate() {
+            lock (this) {
+                interpolateT = Time.time;
+                if (gotData < 2) {
+                    if (valueSet)
+                        return dataValue;
+                    else
+                        return GetDefault();
+                }
+                float t = Mathf.Min((interpolateT - updateT) / (float)(currentT - lastT), clamp);
+                return ValueLerpUnclamped(dataValueLast, dataValue, t);
+            }
+        }
+    }
+    
+    public class InterpolatedColor : InterpolatedValue<Color> {
+        public override Color ValueLerpUnclamped(Color a, Color b, float t) {
+            return Color.LerpUnclamped(a, b, t);
+        }
+        public override Color GetBuiltinDefault() {
+            return Color.clear;
+        }
+    }
+    
+    public class InterpolatedQuaternion : InterpolatedValue<Quaternion> {
+        public override Quaternion ValueLerpUnclamped(Quaternion a, Quaternion b, float t) {
+            return Quaternion.SlerpUnclamped(a, b, t);
+        }
+        public override Quaternion GetBuiltinDefault() {
+            return Quaternion.identity;
+        }
+    }
+
+    public class InterpolatedVector2 : InterpolatedValue<Vector2> {
+        public override Vector2 ValueLerpUnclamped(Vector2 a, Vector2 b, float t) {
+            return Vector2.LerpUnclamped(a, b, t);
+        }
+        public override Vector2 GetBuiltinDefault() {
+            return Vector2.zero;
+        }
+    }
+
+    public class InterpolatedVector3 : InterpolatedValue<Vector3> {
+        public override Vector3 ValueLerpUnclamped(Vector3 a, Vector3 b, float t) {
+            return Vector3.LerpUnclamped(a, b, t);
+        }
+        public override Vector3 GetBuiltinDefault() {
+            return Vector3.zero;
+        }
+    }
+
+    public class InterpolatedVector4 : InterpolatedValue<Vector4> {
+        public override Vector4 ValueLerpUnclamped(Vector4 a, Vector4 b, float t) {
+            return Vector4.LerpUnclamped(a, b, t);
+        }
+        public override Vector4 GetBuiltinDefault() {
+            return Vector4.zero;
+        }
+    }
+
+    public class InterpolatedFloat : InterpolatedValue<float> {
+        public override float ValueLerpUnclamped(float a, float b, float t) {
+            return Mathf.LerpUnclamped(a, b, t);
+        }
+        public override float GetBuiltinDefault() {
+            return 0f;
+        }
+    }
+}

Unity/OpenSeeInterpolation.cs.meta

@@ -0,0 +1,11 @@
+fileFormatVersion: 2
+guid: d61e11f1b6900a7419383be1377f191b
+MonoImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  defaultReferences: []
+  executionOrder: 0
+  icon: {instanceID: 0}
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/OpenSeeKinematicInterpolation.cs

@@ -0,0 +1,108 @@
+using System;
+using System.Collections;
+using System.Collections.Generic;
+using UnityEngine;
+
+namespace OpenSee {
+
+public class OpenSeeKinematicInterpolation : MonoBehaviour
+{
+    public float interpolateClamp = 1f;
+    
+    private Vector3 lastPos;
+    private Vector3 lastPosV;
+    private double lastT;
+    
+    private Vector3 currentPos;
+    private Vector3 currentPosV;
+    private double currentT = 0f;
+    
+    private Vector3 lastRot = Vector3.zero;
+    private Vector3 lastRotV;
+    
+    private Vector3 currentRot;
+    private Vector3 currentRotV;
+    
+    private float dT;
+    private Vector3 coeffPosA;
+    private Vector3 coeffPosB;
+    private Vector3 coeffRotA;
+    private Vector3 coeffRotB;
+    
+    private double updateT;
+    private double interpolateT;
+    private int ready = 0;
+    
+    public void UpdateKI(double newT, Vector3 newPos, Quaternion newRot) {
+        Interpolate();
+        if (newT - currentT > 3f)
+            ready = 0;
+
+        currentPos = transform.localPosition;
+        currentRot = OpenSeeIKTarget.PrettyEuler(currentRot, transform.localRotation.eulerAngles);
+        float dTNow = (float)(interpolateT - updateT);
+        if (interpolateT - updateT > 0 || ready < 1) {
+            currentPosV = (currentPos - lastPos) / dT;
+            currentRotV = (currentRot - lastRot) / dT;
+        } else {
+            currentPosV = Vector3.zero;
+            currentRotV = Vector3.zero;
+        }
+
+        lastPos = currentPos;
+        lastPosV = currentPosV;
+        lastRot = currentRot;
+        lastRotV = currentRotV;
+        lastT = currentT;
+        
+        currentPos = newPos;
+        currentRot = OpenSeeIKTarget.PrettyEuler(lastRot, newRot.eulerAngles);
+        currentT = newT;
+        dT = (float)(currentT - lastT);
+        currentPosV = (currentPos - lastPos) / dT;
+        currentRotV = (currentRot - lastRot) / dT;
+
+        float dTPow2 = Mathf.Pow(dT, 2);
+        float dTPow3 = Mathf.Pow(dT, 3);
+        coeffPosA = (- (2f*(currentPosV * dT + 2f * dT * lastPosV - 3f * currentPos + 3f * lastPos)) / dTPow2) / 2f;
+        coeffPosB = ((6f*(currentPosV * dT + dT * lastPosV - 2f * currentPos + 2f * lastPos)) / dTPow3)/6f;
+        coeffRotA = (- (2f*(currentRotV * dT + 2f * dT * lastRotV - 3f * currentRot + 3f * lastRot)) / dTPow2) / 2f;
+        coeffRotB = ((6f*(currentRotV * dT + dT * lastRotV - 2f * currentRot + 2f * lastRot)) / dTPow3)/6f;
+        
+        updateT = Time.time;
+        if (ready <= 2)
+            ready++;
+    }
+    
+    float InterpolateValue(float dTNow, float dTNowPow2, float dTNowPow3, float a, float b, float x1, float x2, float v1, float v2) {
+        return x1 + v1 * dTNow + dTNowPow2 * a + dTNowPow3 * b;
+    }
+    
+    void Interpolate() {
+        interpolateT = Time.time;
+        if (interpolateT == updateT || ready <= 2)
+            return;
+        float dTNow = (float)(interpolateT - updateT);
+        if (dTNow >= dT) {
+            transform.localPosition = Vector3.LerpUnclamped(lastPos, currentPos, Mathf.Min((float)(dTNow / dT), interpolateClamp));
+            transform.localEulerAngles = Vector3.LerpUnclamped(lastRot, currentRot, Mathf.Min((float)(dTNow / dT), interpolateClamp));
+        } else {
+            float dTNowPow2 = Mathf.Pow(dTNow, 2);
+            float dTNowPow3 = Mathf.Pow(dTNow, 3);
+            float x = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffPosA.x, coeffPosB.x, lastPos.x, currentPos.x, lastPosV.x, currentPosV.x);
+            float y = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffPosA.y, coeffPosB.y, lastPos.y, currentPos.y, lastPosV.y, currentPosV.y);
+            float z = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffPosA.z, coeffPosB.z, lastPos.z, currentPos.z, lastPosV.z, currentPosV.z);
+            transform.localPosition = new Vector3(x, y, z);
+            x = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffRotA.x, coeffRotB.x, lastRot.x, currentRot.x, lastRotV.x, currentRotV.x);
+            y = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffRotA.y, coeffRotB.y, lastRot.y, currentRot.y, lastRotV.y, currentRotV.y);
+            z = InterpolateValue(dTNow, dTNowPow2, dTNowPow3, coeffRotA.z, coeffRotB.z, lastRot.z, currentRot.z, lastRotV.z, currentRotV.z);
+            transform.localEulerAngles = new Vector3(x, y, z);
+        }
+    }
+    
+    public void Update() {
+        Interpolate();
+    }
+}
+
+}

Unity/OpenSeeKinematicInterpolation.cs.meta

@@ -0,0 +1,11 @@
+fileFormatVersion: 2
+guid: a4112b4be90fc9e4baee9eade65378c3
+MonoImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  defaultReferences: []
+  executionOrder: -1
+  icon: {instanceID: 0}
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/OpenSeeLauncher.cs

@@ -51,14 +51,18 @@ public class OpenSeeLauncher : MonoBehaviour {
     public bool autoListCameras = false;
     [Tooltip("If this is enabled, no output from the tracker will be printed to the debug console. This cannot be changed while the tracker is running.")]
     public bool dontPrint = true;
+    [Tooltip("If enabled, the next available port within 500 starting after the one given on the OpenSee component will be used.")]
+    public bool dynamicPort = false;
     [Tooltip("The path to \"facetracker.exe\".")]
     public string exePath = "facetracker.exe";
     [Tooltip("The path, where the .onnx model files can be found")]
     public string modelPath = "models\\";
     [Tooltip("Additional options that should be passed to the face tracker.")]
-    public List<string> commandlineOptions = new List<string>(new string[] { "--silent", "1", "--max-threads", "4" });
+    public List<string> commandlineOptions = new List<string>(new string[] { "--silent", "1", "--max-threads", "1" });
     [Tooltip("This string will be appended at the end of the tracker options without quoting.")]
     public string extraOptions = "";
+    [Tooltip("If enabled, the tracker commandline option string will be logged with Debug.Log.")]
+    public bool logCommandline = false;
     [Tooltip("IL2CPP doesn't support Proocess.Start. When this is enabled, OpenSeeLauncher will create the tracking process by calling the necessary API functions directly through the DLLs. In this case, reading the standard output from the process will not be supported and it will instead be send to the logfiles set in pinvokeStdOut and pinvokeStdErr in the persistent data directory. It will also retrieve the camera list directly through the libdshowcapture DLLs, so make sure it is part of your Unity project.")]
     public bool usePinvoke = false;
     [Tooltip("When this is enabled, even if usePinvoke is disabled, the camera list will be retrieved through libdshowcapture DLLs directly, which can be faster. Make sure the DLLs are in your Unity project.")]
@@ -80,6 +84,30 @@ public class OpenSeeLauncher : MonoBehaviour {
     private System.IntPtr processStdErr = System.IntPtr.Zero;
     private Job job = null;
     
+    public void UnsetOption(string name, bool hasArgument) {
+        int count = 1;
+        if (hasArgument)
+            count++;
+        int found = -1;
+        int i = 0;
+        foreach (string arg in commandlineOptions) {
+            if (arg == name) {
+                found = i;
+                break;
+            }
+            i++;
+        }
+        if (found > -1)
+            commandlineOptions.RemoveRange(found, count);
+    }
+    
+    public void SetOption(string name, string argument) {
+        UnsetOption(name, argument != null);
+        commandlineOptions.Add(name);
+        if (argument != null)
+            commandlineOptions.Add(argument);
+    }
+
     private bool CheckSetup(bool requireTarget) {
         if (openSeeTarget == null)
             openSeeTarget = GetComponent<OpenSee>();
@@ -195,6 +223,48 @@ public class OpenSeeLauncher : MonoBehaviour {
         return cameras;
     }
     
+    public float[] Benchmark(int threads) {
+        if (!CheckSetup(false))
+            return null;
+
+        StringBuilder stringBuilder;
+        ProcessStartInfo processStartInfo;
+        Process process;
+
+        stringBuilder = new StringBuilder();
+
+        processStartInfo = new ProcessStartInfo();
+        processStartInfo.CreateNoWindow = true;
+        processStartInfo.RedirectStandardOutput = true;
+        processStartInfo.RedirectStandardInput = true;
+        processStartInfo.RedirectStandardError = true;
+        processStartInfo.UseShellExecute = false;
+        processStartInfo.FileName = exePath;
+        processStartInfo.Arguments = "--benchmark 1 --priority 4 --max-threads " + threads.ToString();
+        
+        process = new Process();
+        process.StartInfo = processStartInfo;
+        process.EnableRaisingEvents = true;
+        process.OutputDataReceived += new DataReceivedEventHandler(delegate(object sender, DataReceivedEventArgs e) {
+            stringBuilder.Append(e.Data);
+            stringBuilder.Append("\n");
+        });
+        process.Start();
+        job.AddProcess(process.Handle);
+        process.BeginOutputReadLine();
+        process.WaitForExit();
+        process.CancelOutputRead();
+        
+        string[] lines = stringBuilder.ToString().Split(new[] { "\r\n", "\r", "\n" }, StringSplitOptions.RemoveEmptyEntries);
+        float[] results = new float[lines.Length];
+        
+        for (int i = 0; i < lines.Length; i++) {
+            results[i] = Single.Parse(lines[i]);
+        }
+        
+        return results;
+    }
+    
     // From: http://csharptest.net/529/how-to-correctly-escape-command-line-arguments-in-c/index.html
     private static string EscapeArguments(params string[] args)
     {
@@ -227,6 +297,27 @@ public class OpenSeeLauncher : MonoBehaviour {
         List<string> arguments = new List<string>();
         arguments.Add("--ip");
         arguments.Add(ip);
+        
+        if (dynamicPort && !openSeeTarget.listening) {
+            System.Net.IPEndPoint[] inUse = System.Net.NetworkInformation.IPGlobalProperties.GetIPGlobalProperties().GetActiveUdpListeners();
+            Array.Sort(inUse, delegate(System.Net.IPEndPoint a, System.Net.IPEndPoint b) { return a.Port.CompareTo(b.Port); });
+            int port = openSeeTarget.listenPort + 1;
+            for (int i = 0; i < inUse.Length; i++) {
+                if (inUse[i].Port < port)
+                    continue;
+                if (inUse[i].Port == port) {
+                    if (port <= openSeeTarget.listenPort + 500 && port <= IPEndPoint.MaxPort) {
+                        port++;
+                        continue;
+                    } else {
+                        return false;
+                    }
+                }
+                break;
+            }
+            openSeeTarget.listenPort = port;
+        }
+
         arguments.Add("--port");
         arguments.Add(openSeeTarget.listenPort.ToString());
         arguments.Add("--model-dir");
@@ -244,6 +335,9 @@ public class OpenSeeLauncher : MonoBehaviour {
         if (extraOptions != "")
             argumentString = argumentString + " " + extraOptions;
         
+        if (logCommandline)
+            UnityEngine.Debug.Log("Starting tracker: " + argumentString);
+
         StopTracker();
 
         if (!usePinvoke) {

Unity/OpenSeeShowPoints.cs

@@ -7,6 +7,7 @@ namespace OpenSee {
 public class OpenSeeShowPoints : MonoBehaviour {
     public OpenSee openSee = null;
     public int faceId = 0;
+    public bool only30Points = false;
     public bool show3DPoints = true;
     public bool applyTranslation = false;
     public bool applyRotation = false;
@@ -17,6 +18,7 @@ public class OpenSeeShowPoints : MonoBehaviour {
     public bool showLines = false;
     public float lineWidth = 0.01f;
     public Material lineMaterial;
+    public bool receiveShadows = false;
     
     private OpenSee.OpenSeeData openSeeData;
     private GameObject[] gameObjects;
@@ -25,7 +27,10 @@ public class OpenSeeShowPoints : MonoBehaviour {
     private double updated = 0.0;
     private int total = 70;
     
-    private int[] lines = new int[]{/* Contour */ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, -1, /* Eyebrows */ 18, 19, 20, 21, -1, 23, 24, 25, 26, -1, 28, 29, 30, 33, 32, 33, 34, 35, -1, /* Eye */ 37, 38, 39, 40, 41, 36, /* Eye */ 43, 44, 45, 46, 47, 42, /* Mouth */ 49, 50, 51, 52, 62, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 58, 58, 62};
+    private int[] lines = new int[]{/* Contour */ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, -1, /* Eyebrows */ 18, 19, 20, 21, -1, 23, 24, 25, 26, -1, /* Nose */ 28, 29, 30, 33, 32, 33, 34, 35, -1, /* Eye */ 37, 38, 39, 40, 41, 36, /* Eye */ 43, 44, 45, 46, 47, 42, /* Mouth */ 49, 50, 51, 52, 62, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 58, 58, 62};
+    
+    private HashSet<int> pt30Set = new HashSet<int> { 0, 2, 5, 8, 11, 14, 16, 17, 19, 21, 22, 24, 26, 27, 30, 33, 36, 37, 39, 40, 42, 43, 45, 46, 50, 55, 58, 60, 62, 64 };
+    private int[] linesPt30Set = new int[]{/* Contour */ 2, -1, 5, -1, -1, 8, -1, -1, 11, -1, -1, 14, -1, -1, 16, -1, -1, /* Eyebrows */ 19, -1, 21, -1, -1, 24, -1, 26, -1, -1, /* Nose */ 30, -1, -1, 33, -1, -1, -1, -1, -1, /* Eye */ 37, 39, -1, 40, 36, -1, /* Eye */ 43, 45, -1, 46, 42, -1, /* Mouth */ -1, -1, 62, -1, -1, -1, -1, 58, -1, -1, 60, -1, 62, -1, 64, -1, 58, -1, 58, 62};
 
 	void Start () {
         if (openSee == null) {
@@ -42,6 +47,7 @@ public class OpenSeeShowPoints : MonoBehaviour {
             gameObjects[i].name = "Point " + (i + 1);
             if (material != null)
                 gameObjects[i].GetComponent<Renderer>().material = material;
+            gameObjects[i].GetComponent<Renderer>().receiveShadows = receiveShadows;
             gameObjects[i].layer = gameObject.layer;
             gameObjects[i].transform.SetParent(transform);
             gameObjects[i].transform.localScale = new Vector3(0.025f, 0.025f, 0.025f);
@@ -49,22 +55,31 @@ public class OpenSeeShowPoints : MonoBehaviour {
                 GameObject cylinder = GameObject.CreatePrimitive(PrimitiveType.Cylinder);
                 if (material != null)
                     cylinder.GetComponent<Renderer>().material = material;
+                cylinder.GetComponent<Renderer>().receiveShadows = receiveShadows;
                 cylinder.layer = gameObject.layer;
                 cylinder.transform.SetParent(gameObjects[i].transform);
                 cylinder.transform.localEulerAngles = new Vector3(90f, 0f, 0f);
                 cylinder.transform.localPosition = new Vector3(0f, 0f, -4f);
                 cylinder.transform.localScale = new Vector3(1f, 4f, 1f);
             }
+            if (only30Points && i < 66 && !pt30Set.Contains(i))
+                gameObjects[i].SetActive(false);
         }
         for (int i = 0; i < 68; i++) {
                 if (i == 66) {
                     GameObject lineGameObject = new GameObject("LineGameObject");
-                    lineGameObject.transform.SetParent(gameObjects[48].transform);
+                    if (only30Points)
+                        lineGameObject.transform.SetParent(gameObjects[50].transform);
+                    else
+                        lineGameObject.transform.SetParent(gameObjects[48].transform);
                     lineGameObject.layer = gameObject.layer;
                     lineRenderers[i] = lineGameObject.AddComponent(typeof(LineRenderer)) as LineRenderer;
                 } else if (i == 67) {
                     GameObject lineGameObject = new GameObject("LineGameObject");
-                    lineGameObject.transform.SetParent(gameObjects[53].transform);
+                    if (only30Points)
+                        lineGameObject.transform.SetParent(gameObjects[55].transform);
+                    else
+                        lineGameObject.transform.SetParent(gameObjects[53].transform);
                     lineGameObject.layer = gameObject.layer;
                     lineRenderers[i] = lineGameObject.AddComponent(typeof(LineRenderer)) as LineRenderer;
                 }
@@ -73,6 +88,7 @@ public class OpenSeeShowPoints : MonoBehaviour {
                 lineRenderers[i].useWorldSpace = true;
                 lineRenderers[i].generateLightingData = true;
                 lineRenderers[i].material = lineMaterial;
+                lineRenderers[i].receiveShadows = receiveShadows;
                 lineRenderers[i].widthMultiplier = lineWidth;
                 lineRenderers[i].positionCount = 2;
         }
@@ -161,6 +177,7 @@ public class OpenSeeShowPoints : MonoBehaviour {
 
             for (int i = 0; i < 66; i++) {
                 Renderer renderer = gameObjects[i].GetComponent<Renderer>();
+                renderer.receiveShadows = receiveShadows;
                 renderer.material.SetColor("_Color", Color.Lerp(Color.red, Color.green, openSeeData.confidence[i]));
                 float x = openSeeData.points[i].x;
                 float y = -openSeeData.points[i].y;
@@ -172,21 +189,29 @@ public class OpenSeeShowPoints : MonoBehaviour {
             }
         }
         for (int i = 0; i < 68; i++) {
-            if (lines[i] == -1)
+            if ((!only30Points && lines[i] == -1) || (only30Points && linesPt30Set[i] == -1))
                 continue;
             if (!showLines || lineMaterial == null) {
                 lineRenderers[i].enabled = false;
             } else {
                 int a = i;
                 int b = lines[i];
-                if (i == 66)
+                if (only30Points) {
-                    a = 48;
+                    b = linesPt30Set[i];
-                if (i == 67)
+                    if (i == 66)
-                    a = 53;
+                        a = 50;
+                    if (i == 67)
+                        a = 55;
+                } else {
+                    if (i == 66)
+                        a = 48;
+                    if (i == 67)
+                        a = 53;
+                }
                 Color color = Color.Lerp(Color.red, Color.green, Mathf.Lerp(0.5f, openSeeData.confidence[a], openSeeData.confidence[b]));
                 lineRenderers[i].enabled = true;
                 lineRenderers[i].widthMultiplier = lineWidth;
-                //lineRenderers[i].material = lineMaterial;
+                lineRenderers[i].receiveShadows = receiveShadows;
                 lineRenderers[i].material.SetColor("_Color", color);
                 lineRenderers[i].startColor = color;
                 lineRenderers[i].endColor = color;

Unity/Plugins/SVMModel.cs

@@ -6,34 +6,61 @@ using System.Runtime.InteropServices;
 using System.Runtime.Serialization;
 using System.Runtime.Serialization.Formatters.Binary;
 using System.Text;
+using UnityEngine;
 
 namespace OpenSee {
 
 public class SVMModel {
 	#region DllImport
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, EntryPoint = "trainModel")]
-	private static extern System.IntPtr trainModel(float[] features, float[] labels, float[] weights, int rows, int cols, int classes, int probability, float C);
+	private static extern System.IntPtr trainModelDll(float[] features, float[] labels, float[] weights, int rows, int cols, int classes, int probability, float C);
+    protected virtual System.IntPtr trainModel(float[] features, float[] labels, float[] weights, int rows, int cols, int classes, int probability, float C) {
+        Debug.Log("Training LibSVM");
+        return trainModelDll(features, labels, weights, rows, cols, classes, probability, C);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, EntryPoint = "predict")]
-	private static extern void predict(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows);
+	private static extern void predictDll(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows);
+    protected virtual void predict(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows) {
+        predictDll(model, features, predictions, probabilities, rows);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "loadModel")]
-	private static extern System.IntPtr loadModel([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs);
+	private static extern System.IntPtr loadModelDll([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs);
+    protected virtual System.IntPtr loadModel([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs) {
+        return loadModelDll(filename, cols, classes, means, sdevs);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "loadModelString")]
-	private static extern System.IntPtr loadModelString([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs);
+	private static extern System.IntPtr loadModelStringDll([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs);
+    protected virtual System.IntPtr loadModelString([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs) {
+        Debug.Log("Loading LibSVM");
+        return loadModelStringDll(modelString, cols, classes, means, sdevs);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, EntryPoint = "getScales")]
-	private static extern void getScales(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs);
+	private static extern void getScalesDll(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs);
+    protected virtual void getScales(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs) {
+        getScalesDll(model, means, sdevs);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "saveModel")]
-	private static extern int saveModel(System.IntPtr model, [MarshalAs(UnmanagedType.LPStr)]string filename);
+	private static extern int saveModelDll(System.IntPtr model, [MarshalAs(UnmanagedType.LPStr)]string filename);
+    protected virtual int saveModel(System.IntPtr model, [MarshalAs(UnmanagedType.LPStr)]string filename) {
+        return saveModelDll(model, filename);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "saveModelString")]
-	private static extern string saveModelString(System.IntPtr model);
+	private static extern string saveModelStringDll(System.IntPtr model);
+    protected virtual string saveModelString(System.IntPtr model) {
+        return saveModelStringDll(model);
+    }
 
-	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
+	[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, EntryPoint = "destroyModel")]
-	private static extern void destroyModel(System.IntPtr model);
+	private static extern void destroyModelDll(System.IntPtr model);
+    protected virtual void destroyModel(System.IntPtr model) {
+        destroyModelDll(model);
+    }
 	#endregion
 
     private bool haveModel = false;
@@ -43,51 +70,51 @@ public class SVMModel {
 
     [Serializable]
     private class SVMModelRepresentation {
-        private string modelString;
+        public string modelString;
-        private int cols;
+        public int cols;
-        private int maxClasses;
+        public int maxClasses;
-        private double[] means;
+        public double[] means;
-        private double[] sdevs;
+        public double[] sdevs;
+    }
 
-        static public System.IntPtr LoadSerialized(byte[] modelBytes, out int cols, out int maxClasses, bool compress) {
+    public System.IntPtr LoadSerialized(byte[] modelBytes, out int cols, out int maxClasses, bool compress) {
-            IFormatter formatter = new BinaryFormatter();
+        IFormatter formatter = new BinaryFormatter();
-            MemoryStream memoryStream = new MemoryStream(modelBytes);
+        MemoryStream memoryStream = new MemoryStream(modelBytes);
-            memoryStream.Position = 0;
+        memoryStream.Position = 0;
-            SVMModelRepresentation smr;
+        SVMModelRepresentation smr;
-            if (compress) {
+        if (compress) {
-                using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Decompress)) {
+            using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Decompress)) {
-                    smr = formatter.Deserialize(gzipStream) as SVMModelRepresentation;
+                smr = formatter.Deserialize(gzipStream) as SVMModelRepresentation;
-                }
-            } else {
-                smr = formatter.Deserialize(memoryStream) as SVMModelRepresentation;
             }
-            cols = smr.cols;
+        } else {
-            maxClasses = smr.maxClasses;
+            smr = formatter.Deserialize(memoryStream) as SVMModelRepresentation;
-            System.IntPtr model = loadModelString(smr.modelString, smr.cols, smr.maxClasses, smr.means, smr.sdevs);
-            return model;
         }
+        cols = smr.cols;
+        maxClasses = smr.maxClasses;
+        System.IntPtr model = loadModelString(smr.modelString, smr.cols, smr.maxClasses, smr.means, smr.sdevs);
+        return model;
+    }
 
-        static public byte[] ToSerialized(System.IntPtr model, int cols, int maxClasses, bool compress) {
+    public byte[] ToSerialized(System.IntPtr model, int cols, int maxClasses, bool compress) {
-            SVMModelRepresentation smr = new SVMModelRepresentation();
+        SVMModelRepresentation smr = new SVMModelRepresentation();
-            smr.cols = cols;
+        smr.cols = cols;
-            smr.maxClasses = maxClasses;
+        smr.maxClasses = maxClasses;
-            smr.means = new double[cols];
+        smr.means = new double[cols];
-            smr.sdevs = new double[cols];
+        smr.sdevs = new double[cols];
-            getScales(model, smr.means, smr.sdevs);
+        getScales(model, smr.means, smr.sdevs);
-            smr.modelString = saveModelString(model);
+        smr.modelString = saveModelString(model);
 
-            IFormatter formatter = new BinaryFormatter();
+        IFormatter formatter = new BinaryFormatter();
-            MemoryStream memoryStream = new MemoryStream();
+        MemoryStream memoryStream = new MemoryStream();
-            if (compress) {
+        if (compress) {
-                using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Compress)) {
+            using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Compress)) {
-                    formatter.Serialize(gzipStream, smr);
+                formatter.Serialize(gzipStream, smr);
-                    gzipStream.Flush();
+                gzipStream.Flush();
-                }
-            } else {
-                formatter.Serialize(memoryStream, smr);
             }
-            return memoryStream.ToArray();
+        } else {
+            formatter.Serialize(memoryStream, smr);
         }
+        return memoryStream.ToArray();
     }
 
     public void DestroyModel() {
@@ -99,7 +126,7 @@ public class SVMModel {
 
     private void LoadModel(byte[] modelBytes, bool compress) {
         DestroyModel();
-        model = SVMModelRepresentation.LoadSerialized(modelBytes, out cols, out maxClasses, compress);
+        model = LoadSerialized(modelBytes, out cols, out maxClasses, compress);
         haveModel = true;
     }
 
@@ -110,7 +137,7 @@ public class SVMModel {
     private byte[] SaveModel(bool compress) {
         if (!haveModel)
             return null;
-        return SVMModelRepresentation.ToSerialized(model, cols, maxClasses, compress);
+        return ToSerialized(model, cols, maxClasses, compress);
     }
 
     public byte[] SaveModel() {
@@ -143,7 +170,9 @@ public class SVMModel {
 
         maxClasses = max + 1;
         this.cols = cols;
+        float start = Time.realtimeSinceStartup;
         model = trainModel(features, labels, weights, rows, cols, maxClasses, probability, C);
+        Debug.Log("Training took: " + (Time.realtimeSinceStartup - start).ToString());
         haveModel = true;
 
         return true;

Unity/Plugins/ThunderSVMModel.cs

@@ -0,0 +1,71 @@
+using System;
+using System.Collections.Generic;
+using System.IO;
+using System.IO.Compression;
+using System.Runtime.InteropServices;
+using System.Runtime.Serialization;
+using System.Runtime.Serialization.Formatters.Binary;
+using System.Text;
+using OpenSee;
+using UnityEngine;
+
+public class ThunderSVMModel : SVMModel {
+	#region DllImport
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, EntryPoint = "trainModel")]
+	private static extern System.IntPtr trainModelDll(float[] features, float[] labels, float[] weights, int rows, int cols, int classes, int probability, float C);
+    protected override System.IntPtr trainModel(float[] features, float[] labels, float[] weights, int rows, int cols, int classes, int probability, float C) {
+        Debug.Log("Training ThunderSVM");
+        return trainModelDll(features, labels, weights, rows, cols, classes, probability, C);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, EntryPoint = "predict")]
+	private static extern void predictDll(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows);
+    protected override void predict(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows) {
+        predictDll(model, features, predictions, probabilities, rows);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "loadModel")]
+	private static extern System.IntPtr loadModelDll([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs);
+    protected override System.IntPtr loadModel([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs) {
+        return loadModelDll(filename, cols, classes, means, sdevs);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "loadModelString")]
+	private static extern System.IntPtr loadModelStringDll([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs);
+    protected override System.IntPtr loadModelString([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs) {
+        Debug.Log("Loading ThunderSVM");
+        return loadModelStringDll(modelString, cols, classes, means, sdevs);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, EntryPoint = "getScales")]
+	private static extern void getScalesDll(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs);
+    protected override void getScales(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs) {
+        getScalesDll(model, means, sdevs);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "saveModel")]
+	private static extern int saveModelDll(System.IntPtr model, [MarshalAs(UnmanagedType.LPStr)]string filename);
+    protected override int saveModel(System.IntPtr model, [MarshalAs(UnmanagedType.LPStr)]string filename) {
+        return saveModelDll(model, filename);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi, EntryPoint = "saveModelString")]
+	private static extern string saveModelStringDll(System.IntPtr model);
+    protected override string saveModelString(System.IntPtr model) {
+        return saveModelStringDll(model);
+    }
+
+	[DllImport("ThunderSVMWrapper", CallingConvention = CallingConvention.Cdecl, EntryPoint = "destroyModel")]
+	private static extern void destroyModelDll(System.IntPtr model);
+    protected override void destroyModel(System.IntPtr model) {
+        destroyModelDll(model);
+    }
+	#endregion
+    
+    public ThunderSVMModel() {
+    }
+
+    public ThunderSVMModel(byte[] modelBytes) {
+        LoadModel(modelBytes);
+    }
+}

Unity/Plugins/ThunderSVMWrapper.dll.meta

@@ -0,0 +1,63 @@
+fileFormatVersion: 2
+guid: 9c58dd9d1f0cf3140bd004d536712abb
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      : Any
+    second:
+      enabled: 0
+      settings:
+        Exclude Editor: 0
+        Exclude Linux64: 0
+        Exclude OSXUniversal: 0
+        Exclude Win: 1
+        Exclude Win64: 0
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 1
+      settings:
+        CPU: AnyCPU
+        DefaultValueInitialized: true
+        OS: AnyOS
+  - first:
+      Standalone: Linux64
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  - first:
+      Standalone: OSXUniversal
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  - first:
+      Standalone: Win
+    second:
+      enabled: 0
+      settings:
+        CPU: None
+  - first:
+      Standalone: Win64
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/Plugins/dshowcapture_x64.dll.meta

@@ -0,0 +1,27 @@
+fileFormatVersion: 2
+guid: dd0c04eb3445c3048bfaa964c6bcb16f
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 0
+      settings:
+        DefaultValueInitialized: true
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/Plugins/dshowcapture_x86.dll.meta

@@ -0,0 +1,27 @@
+fileFormatVersion: 2
+guid: a504a0f747083b54c99d6e91731ccecc
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 0
+      settings:
+        DefaultValueInitialized: true
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/Plugins/libminibmcapture32.dll.meta

@@ -0,0 +1,27 @@
+fileFormatVersion: 2
+guid: 51dc8f0c65df30d4daa188cdfd1dd220
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 0
+      settings:
+        DefaultValueInitialized: true
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/Plugins/libminibmcapture64.dll.meta

@@ -0,0 +1,27 @@
+fileFormatVersion: 2
+guid: fcfc753ca92169e48aa69c1ecc1b0d64
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 0
+      settings:
+        DefaultValueInitialized: true
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

Unity/Plugins/thundersvm.dll.meta

@@ -0,0 +1,63 @@
+fileFormatVersion: 2
+guid: e549e6cbc6a547748abc7f1b05bfa7b4
+PluginImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  iconMap: {}
+  executionOrder: {}
+  defineConstraints: []
+  isPreloaded: 0
+  isOverridable: 0
+  isExplicitlyReferenced: 0
+  validateReferences: 1
+  platformData:
+  - first:
+      : Any
+    second:
+      enabled: 0
+      settings:
+        Exclude Editor: 0
+        Exclude Linux64: 0
+        Exclude OSXUniversal: 0
+        Exclude Win: 1
+        Exclude Win64: 0
+  - first:
+      Any: 
+    second:
+      enabled: 1
+      settings: {}
+  - first:
+      Editor: Editor
+    second:
+      enabled: 1
+      settings:
+        CPU: AnyCPU
+        DefaultValueInitialized: true
+        OS: AnyOS
+  - first:
+      Standalone: Linux64
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  - first:
+      Standalone: OSXUniversal
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  - first:
+      Standalone: Win
+    second:
+      enabled: 0
+      settings:
+        CPU: None
+  - first:
+      Standalone: Win64
+    second:
+      enabled: 1
+      settings:
+        CPU: x86_64
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

dshowcapture.py

@@ -5,53 +5,95 @@ import numpy as np
 from ctypes import *
 from PIL import Image
 import cv2
+import json
 
 def resolve(name):
     f = os.path.join(os.path.dirname(__file__), name)
     return f
 
 lib = None
+bm_lib = None
+bm_options = None
+bm_enabled = False
 
-def create_frame_buffer(width, height):
+def set_bm_enabled(v):
-    buffer = bytearray(width * height * 4)
+    global bm_enabled
+    bm_enabled = v
+
+def set_options(str):
+    global bm_options
+    bm_options = str
+
+def create_frame_buffer(width, height, factor):
+    buffer = bytearray(width * height * 4 * factor)
     char_array = c_char * len(buffer)
     return char_array.from_buffer(buffer)
 
 class DShowCapture():
     def __init__(self):
         global lib
-        if lib is None:
+        global bm_lib
+        global bm_options
+        global bm_enabled
+        if lib is None or bm_lib is None:
             if platform.architecture()[0] == '32bit':
                 dll_path = resolve(os.path.join("dshowcapture", "dshowcapture_x86.dll"))
                 lib = cdll.LoadLibrary(dll_path)
+                if bm_enabled:
+                    dll_path = resolve(os.path.join("dshowcapture", "libminibmcapture32.dll"))
+                    bm_lib = cdll.LoadLibrary(dll_path)
             else:
                 dll_path = resolve(os.path.join("dshowcapture", "dshowcapture_x64.dll"))
                 lib = cdll.LoadLibrary(dll_path)
+                if bm_enabled:
+                    dll_path = resolve(os.path.join("dshowcapture", "libminibmcapture64.dll"))
+                    bm_lib = cdll.LoadLibrary(dll_path)
+
+            # DirectShow
             lib.create_capture.restype = c_void_p
             lib.get_devices.argtypes = [c_void_p]
             lib.get_device.argtypes = [c_void_p, c_int, c_char_p, c_int]
             lib.capture_device.argtypes = [c_void_p, c_int, c_int, c_int, c_int]
+            lib.capture_device_by_dcap.argtypes = [c_void_p, c_int, c_int, c_int, c_int, c_longlong]
             lib.capture_device_default.argtypes = [c_void_p, c_int]
             lib.get_width.argtypes = [c_void_p]
             lib.get_height.argtypes = [c_void_p]
             lib.get_fps.argtypes = [c_void_p]
             lib.get_flipped.argtypes = [c_void_p]
             lib.get_colorspace.argtypes = [c_void_p]
+            lib.get_colorspace_internal.argtypes = [c_void_p]
             lib.capturing.argtypes = [c_void_p]
             lib.get_frame.argtypes = [c_void_p, c_int, c_char_p, c_int]
             lib.get_size.argtypes = [c_void_p]
             lib.stop_capture.argtypes = [c_void_p]
+            lib.get_json_length.argtypes = [c_void_p]
+            lib.get_json.argtypes = [c_void_p, c_char_p, c_int]
             lib.destroy_capture.argtypes = [c_void_p]
+
+            # Blackmagic
+            if bm_enabled:
+                bm_lib.start_capture_single.argtypes = [c_int, c_int, c_void_p]
+                bm_lib.read_frame_bgra32_blocking.argtypes = [c_char_p, c_int]
+                bm_lib.stop_capture_single.argtypes = []
+                bm_lib.get_json_length.argtypes = []
+                bm_lib.get_json.argtypes = [c_char_p, c_int]
+                if bm_options is not None:
+                    bm_lib.set_options.argtypes = [c_char_p]
+                    bm_lib.set_options(bm_options.encode())
         self.lib = lib
+        if bm_enabled:
+            self.bm_lib = bm_lib
         self.cap = lib.create_capture()
         self.name_buffer = create_string_buffer(255);
         self.buffer = None
         self.have_devices = False
         self.size = None
         self.real_size = None
+        self.info = None
+        self.type = None
 
     def __del__(self):
-        if not self.buffer is None:
+        if self.buffer is not None:
             del self.buffer
         del self.name_buffer
         self.destroy_capture()
@@ -62,20 +104,88 @@ class DShowCapture():
 
     def get_device(self, device_number):
         self.lib.get_device(self.cap, device_number, self.name_buffer, 255)
-        name_str = str(self.name_buffer.value.decode('utf8'))
+        name_str = str(self.name_buffer.value.decode('utf8', 'surrogateescape'))
         return name_str
 
+    def get_info(self):
+        global bm_enabled
+        # DirectShow
+        json_length = self.lib.get_json_length(self.cap);
+        json_buffer = create_string_buffer(json_length)
+        self.lib.get_json(self.cap, json_buffer, json_length);
+        json_text = str(json_buffer.value.decode('utf8', 'surrogateescape'))
+        self.info = json.loads(json_text)
+        for cam in self.info:
+            cam["type"] = "DirectShow"
+            cam["index"] = cam["id"]
+        
+        # Blackmagic
+        if bm_enabled:
+            json_length = self.bm_lib.get_json_length();
+            json_buffer = create_string_buffer(json_length)
+            self.bm_lib.get_json(json_buffer, json_length);
+            json_text = str(json_buffer.value.decode('utf8', 'surrogateescape'))
+            bm_info = json.loads(json_text)
+            dshow_len = len(self.info)
+            for cam in bm_info:
+                cam["type"] = "Blackmagic"
+                cam["index"] = cam["id"] + dshow_len
+            self.info.extend(bm_info)
+
+        return self.info
+
     def capture_device(self, cam, width, height, fps):
         if not self.have_devices:
             self.get_devices()
         ret = self.lib.capture_device(self.cap, cam, width, height, fps) == 1
         if ret:
+            self.type = "DirectShow"
             self.width = self.get_width()
             self.height = self.get_height()
+            self.fps = self.get_fps()
             self.flipped = self.get_flipped()
             self.colorspace = self.get_colorspace()
+            self.colorspace_internal = self.get_colorspace_internal()
             self.size = self.width * self.height * 4
-            self.buffer = create_frame_buffer(self.width, self.height)
+            self.buffer = create_frame_buffer(self.width, self.height, 4)
+        else:
+            self.size = None
+            self.real_size = None
+        return ret;
+
+    def capture_device_by_dcap(self, cam, dcap, width, height, fps):
+        global bm_enabled
+        if not self.have_devices:
+            self.get_devices()
+        if self.info is None:
+            self.get_info()
+        fps = int(10000000 / fps)
+        ret = False
+        if self.info[cam]['type'] == "DirectShow":
+            ret = self.lib.capture_device_by_dcap(self.cap, cam, dcap, width, height, fps) == 1
+        elif bm_enabled and self.info[cam]['type'] == "Blackmagic":
+            ret = self.bm_lib.start_capture_single(self.info[cam]['id'], self.info[cam]['caps'][dcap]['bmModecode'], None) == 1
+        if ret:
+            self.type = self.info[cam]['type']
+            if self.type == "DirectShow":
+                self.width = self.get_width()
+                self.height = self.get_height()
+                self.fps = self.get_fps()
+                self.flipped = self.get_flipped()
+                self.colorspace = self.get_colorspace()
+                self.colorspace_internal = self.get_colorspace_internal()
+                self.size = self.width * self.height * 4 * 4
+                self.buffer = create_frame_buffer(self.width, self.height, 4)
+            elif bm_enabled and self.type == "Blackmagic":
+                self.width = self.info[cam]['caps'][dcap]['minCX']
+                self.height = self.info[cam]['caps'][dcap]['minCY']
+                self.fps = int(10000000 / self.info[cam]['caps'][dcap]['minInterval'])
+                self.flipped = True
+                self.colorspace = 101
+                self.colorspace_internal = self.colorspace
+                self.size = self.width * self.height * 4
+                self.buffer = create_frame_buffer(self.width, self.height, 1)
+                self.real_size = self.size
         else:
             self.size = None
             self.real_size = None
@@ -86,12 +196,15 @@ class DShowCapture():
             self.get_devices()
         ret = self.lib.capture_device_default(self.cap, cam) == 1
         if ret:
+            self.type = "DirectShow"
             self.width = self.get_width()
             self.height = self.get_height()
+            self.fps = self.get_fps()
             self.flipped = self.get_flipped()
             self.colorspace = self.get_colorspace()
-            self.size = self.width * self.height * 4
+            self.colorspace_internal = self.get_colorspace_internal()
-            self.buffer = create_frame_buffer(self.width, self.height)
+            self.size = self.width * self.height * 4 * 4
+            self.buffer = create_frame_buffer(self.width, self.height, 4)
         else:
             self.size = None
             self.real_size = None
@@ -110,52 +223,78 @@ class DShowCapture():
         return self.lib.get_flipped(self.cap) != 0
 
     def get_colorspace(self):
-        return self.lib.get_colorspace(self.cap)
+        colorspace = self.lib.get_colorspace(self.cap)
+        if colorspace == 0:
+            colorspace = self.lib.get_colorspace_internal(self.cap)
+        return colorspace
+
+    def get_colorspace_internal(self):
+        return self.lib.get_colorspace_internal(self.cap)
 
     def capturing(self):
+        if self.type == "Blackmagic":
+            return 1
         return self.lib.capturing(self.cap) == 1
 
     def get_frame(self, timeout):
+        global bm_enabled
         if self.size is None:
             return None
-        if self.lib.get_frame(self.cap, timeout, self.buffer, self.size) == 0:
+        if self.type == "DirectShow":
+            self.real_size = self.lib.get_frame(self.cap, timeout, self.buffer, self.size)
+        elif bm_enabled and self.type == "Blackmagic":
+            self.bm_lib.read_frame_bgra32_blocking(self.buffer, self.size)
+        else:
             return None
-        if self.real_size is None:
-            self.real_size = self.lib.get_size(self.cap)
         img = np.frombuffer(self.buffer, dtype=np.uint8)[0:self.real_size]
         if self.colorspace in [100, 101]:
-            img = cv2.cvtColor(img.reshape((self.height,self.width,4)), cv2.COLOR_BGRA2BGR);
+            if self.real_size == self.height * self.width * 4:
+                img = cv2.cvtColor(img.reshape((self.height,self.width,4)), cv2.COLOR_BGRA2BGR)
+            elif self.real_size == self.height * self.width * 3:
+                img = img.reshape((self.height,self.width,3))
+            else:
+                return None
             if not self.flipped:
                 img = cv2.flip(img, 0)
             return img
-        elif self.colorspace == 200:
+        elif self.colorspace == 200 and self.real_size == (3 * self.height // 2) * self.width:
-            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_I420);
+            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_I420)
-        elif self.colorspace == 201:
+        elif self.colorspace == 201 and self.real_size == (3 * self.height // 2) * self.width:
-            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_NV12);
+            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_NV12)
-        elif self.colorspace == 202:
+        elif self.colorspace == 202 and self.real_size == (3 * self.height // 2) * self.width:
-            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_YV12);
+            img = cv2.cvtColor(img.reshape((3 * self.height // 2,self.width,1)), cv2.COLOR_YUV2BGR_YV12)
-        elif self.colorspace == 203:
+        elif self.colorspace == 203 and self.real_size == self.height * self.width:
-            img = cv2.cvtColor(img.reshape((self.height,self.width,1)), cv2.COLOR_GRAY2BGR);
+            img = cv2.cvtColor(img.reshape((self.height,self.width,1)), cv2.COLOR_GRAY2BGR)
-        elif self.colorspace == 300:
+        elif self.colorspace == 300 and self.real_size == self.height * self.width * 2:
-            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_YVYU);
+            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_YVYU)
-        elif self.colorspace == 301:
+        elif self.colorspace == 301 and self.real_size == self.height * self.width * 2:
-            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_YUY2);
+            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_YUY2)
-        elif self.colorspace == 302:
+        elif self.colorspace == 302 and self.real_size == self.height * self.width * 2:
-            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_UYVY);
+            img = cv2.cvtColor(img.reshape((self.height,self.width,2)), cv2.COLOR_YUV2BGR_UYVY)
         elif self.colorspace == 303:
             return None
+        elif self.colorspace == 400 and self.real_size > 4:
+            img = cv2.imdecode(img, cv2.IMREAD_COLOR);
+        else:
+            return None
         if self.flipped:
             img = cv2.flip(img, 0)
         return img
 
     def stop_capture(self):
+        global bm_enabled
         self.size = None
         self.real_size = None
-        return self.lib.stop_capture(self.cap)
+        if self.type == "DirectShow":
+            return self.lib.stop_capture(self.cap)
+        elif bm_enabled and self.type == "Blackmagic":
+            return self.bm_lib.stop_capture_single() != 0
+        return True
 
     def destroy_capture(self):
         if self.cap is None:
             return 0
+        self.stop_capture()
         ret = self.lib.destroy_capture(self.cap)
         self.cap = None
         self.size = None
@@ -186,11 +325,11 @@ if __name__ == "__main__":
     width = cap.get_width()
     height = cap.get_height()
     flipped = cap.get_flipped()
-    print(f"Width: {width} Height: {height} FPS: {cap.get_fps()} Flipped: {flipped} Colorspace: {cap.get_colorspace()}")
+    print(f"Width: {width} Height: {height} FPS: {cap.get_fps()} Flipped: {flipped} Colorspace: {cap.get_colorspace()} Internal: {cap.get_colorspace_internal()}")
     
     while True:
         img = cap.get_frame(1000)
-        if not img is None:
+        if img is not None:
             cv2.imshow("DShowCapture", img)
             if cv2.waitKey(1) & 0xFF == ord('q'):
                 sys.exit(0)

facetracker.py

@@ -2,21 +2,30 @@ import copy
 import os
 import sys
 import argparse
+import traceback
+import gc
 
 parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
 parser.add_argument("-i", "--ip", help="Set IP address for sending tracking data", default="127.0.0.1")
 parser.add_argument("-p", "--port", type=int, help="Set port for sending tracking data", default=11573)
 if os.name == 'nt':
-    parser.add_argument("-l", "--list-cameras", type=int, help="Set this to 1 to list the available cameras and quit, set this to a number higher than 1 to output only the names", default=0)
+    parser.add_argument("-l", "--list-cameras", type=int, help="Set this to 1 to list the available cameras and quit, set this to 2 or higher to output only the names", default=0)
+    parser.add_argument("-a", "--list-dcaps", type=int, help="Set this to -1 to list all cameras and their available capabilities, set this to a camera id to list that camera's capabilities", default=None)
     parser.add_argument("-W", "--width", type=int, help="Set camera and raw RGB width", default=640)
     parser.add_argument("-H", "--height", type=int, help="Set camera and raw RGB height", default=360)
-    parser.add_argument("-F", "--fps", type=int, help="Set camera frames per second", default=24)
+    parser.add_argument("-D", "--dcap", type=int, help="Set which device capability line to use or -1 to use the default camera settings (FPS still need to be set separately)", default=None)
+    parser.add_argument("-B", "--blackmagic", type=int, help="When set to 1, special support for Blackmagic devices is enabled", default=0)
 else:
     parser.add_argument("-W", "--width", type=int, help="Set raw RGB width", default=640)
     parser.add_argument("-H", "--height", type=int, help="Set raw RGB height", default=360)
+if sys.platform == 'linux':
+    parser.add_argument("--dformat", type=str, help="Set which device format to use (MJPG, YUYV, RGB3, ...)", default=None)
+parser.add_argument("-F", "--fps", type=int, help="Set camera frames per second", default=24)
 parser.add_argument("-c", "--capture", help="Set camera ID (0, 1...) or video file", default="0")
+parser.add_argument("-M", "--mirror-input", action="store_true", help="Process a mirror image of the input video")
 parser.add_argument("-m", "--max-threads", type=int, help="Set the maximum number of threads", default=1)
-parser.add_argument("-t", "--threshold", type=float, help="Set minimum confidence threshold for face detection", default=0.65)
+parser.add_argument("-t", "--threshold", type=float, help="Set minimum confidence threshold for face tracking", default=None)
+parser.add_argument("-d", "--detection-threshold", type=float, help="Set minimum confidence threshold for face detection", default=0.6)
 parser.add_argument("-v", "--visualize", type=int, help="Set this to 1 to visualize the tracking, to 2 to also show face ids, to 3 to add confidence values or to 4 to add numbers to the point display", default=0)
 parser.add_argument("-P", "--pnp-points", type=int, help="Set this to 1 to add the 3D fitting points to the visualization", default=0)
 parser.add_argument("-s", "--silent", type=int, help="Set this to 1 to prevent text output on the console", default=0)
@@ -26,33 +35,92 @@ parser.add_argument("--scan-every", type=int, help="Set after how many frames a
 parser.add_argument("--discard-after", type=int, help="Set the how long the tracker should keep looking for lost faces", default=10)
 parser.add_argument("--max-feature-updates", type=int, help="This is the number of seconds after which feature min/max/medium values will no longer be updated once a face has been detected.", default=900)
 parser.add_argument("--no-3d-adapt", type=int, help="When set to 1, the 3D face model will not be adapted to increase the fit", default=1)
+parser.add_argument("--try-hard", type=int, help="When set to 1, the tracker will try harder to find a face", default=0)
 parser.add_argument("--video-out", help="Set this to the filename of an AVI file to save the tracking visualization as a video", default=None)
+parser.add_argument("--video-scale", type=int, help="This is a resolution scale factor applied to the saved AVI file", default=1, choices=[1,2,3,4])
+parser.add_argument("--video-fps", type=float, help="This sets the frame rate of the output AVI file", default=24)
 parser.add_argument("--raw-rgb", type=int, help="When this is set, raw RGB frames of the size given with \"-W\" and \"-H\" are read from standard input instead of reading a video", default=0)
 parser.add_argument("--log-data", help="You can set a filename to which tracking data will be logged here", default="")
-parser.add_argument("--model", type=int, help="This can be used to select the tracking model. Higher numbers are models with better tracking quality, but slower speed. Models 1 and 0 tend to be too rigid for expression and blink detection.", default=3, choices=[0, 1, 2, 3])
+parser.add_argument("--log-output", help="You can set a filename to console output will be logged here", default="")
+parser.add_argument("--model", type=int, help="This can be used to select the tracking model. Higher numbers are models with better tracking quality, but slower speed, except for model 4, which is wink optimized. Models 1 and 0 tend to be too rigid for expression and blink detection. Model -2 is roughly equivalent to model 1, but faster. Model -3 is between models 0 and -1.", default=3, choices=[-3, -2, -1, 0, 1, 2, 3, 4])
 parser.add_argument("--model-dir", help="This can be used to specify the path to the directory containing the .onnx model files", default=None)
-parser.add_argument("--gaze-tracking", type=int, help="When set to 1, experimental blink detection and gaze tracking are enabled, which makes things slightly slower", default=1)
+parser.add_argument("--gaze-tracking", type=int, help="When set to 1, gaze tracking is enabled, which makes things slightly slower", default=1)
 parser.add_argument("--face-id-offset", type=int, help="When set, this offset is added to all face ids, which can be useful for mixing tracking data from multiple network sources", default=0)
 parser.add_argument("--repeat-video", type=int, help="When set to 1 and a video file was specified with -c, the tracker will loop the video until interrupted", default=0)
 parser.add_argument("--dump-points", type=str, help="When set to a filename, the current face 3D points are made symmetric and dumped to the given file when quitting the visualization with the \"q\" key", default="")
+parser.add_argument("--benchmark", type=int, help="When set to 1, the different tracking models are benchmarked, starting with the best and ending with the fastest and with gaze tracking disabled for models with negative IDs", default=0)
 if os.name == 'nt':
     parser.add_argument("--use-dshowcapture", type=int, help="When set to 1, libdshowcapture will be used for video input instead of OpenCV", default=1)
+    parser.add_argument("--blackmagic-options", type=str, help="When set, this additional option string is passed to the blackmagic capture library", default=None)
+    parser.add_argument("--priority", type=int, help="When set, the process priority will be changed", default=None, choices=[0, 1, 2, 3, 4, 5])
 args = parser.parse_args()
 
-os.environ["OMP_NUM_THREADS"] = "1"
+os.environ["OMP_NUM_THREADS"] = str(args.max_threads)
 
-if os.name == 'nt' and args.list_cameras > 0:
+class OutputLog(object):
+    def __init__(self, fh, output):
+        self.fh = fh
+        self.output = output
+    def write(self, buf):
+        if self.fh is not None:
+            self.fh.write(buf)
+        self.output.write(buf)
+        self.flush()
+    def flush(self):
+        if self.fh is not None:
+            self.fh.flush()
+        self.output.flush()
+output_logfile = None
+if args.log_output != "":
+    output_logfile = open(args.log_output, "w")
+sys.stdout = OutputLog(output_logfile, sys.stdout)
+sys.stderr = OutputLog(output_logfile, sys.stderr)
+
+if os.name == 'nt':
     import dshowcapture
+    if args.blackmagic == 1:
+        dshowcapture.set_bm_enabled(True)
+    if args.blackmagic_options is not None:
+        dshowcapture.set_options(args.blackmagic_options)
+    if args.priority is not None:
+        import psutil
+        classes = [psutil.IDLE_PRIORITY_CLASS, psutil.BELOW_NORMAL_PRIORITY_CLASS, psutil.NORMAL_PRIORITY_CLASS, psutil.ABOVE_NORMAL_PRIORITY_CLASS, psutil.HIGH_PRIORITY_CLASS, psutil.REALTIME_PRIORITY_CLASS]
+        p = psutil.Process(os.getpid())
+        p.nice(classes[args.priority])
+
+if os.name == 'nt' and (args.list_cameras > 0 or args.list_dcaps is not None):
     cap = dshowcapture.DShowCapture()
-    camera_count = cap.get_devices()
+    info = cap.get_info()
-    if args.list_cameras == 1:
+    unit = 10000000.;
-        print("Available cameras:")
+    if args.list_dcaps is not None:
-    for i in range(camera_count):
+        formats = {0: "Any", 1: "Unknown", 100: "ARGB", 101: "XRGB", 200: "I420", 201: "NV12", 202: "YV12", 203: "Y800", 300: "YVYU", 301: "YUY2", 302: "UYVY", 303: "HDYC (Unsupported)", 400: "MJPEG", 401: "H264" }
-        camera_name = cap.get_device(i)
+        for cam in info:
+            if args.list_dcaps == -1:
+                type = ""
+                if cam['type'] == "Blackmagic":
+                    type = "Blackmagic: "
+                print(f"{cam['index']}: {type}{cam['name']}")
+            if args.list_dcaps != -1 and args.list_dcaps != cam['index']:
+                continue
+            for caps in cam['caps']:
+                format = caps['format']
+                if caps['format'] in formats:
+                    format = formats[caps['format']]
+                if caps['minCX'] == caps['maxCX'] and caps['minCY'] == caps['maxCY']:
+                    print(f"    {caps['id']}: Resolution: {caps['minCX']}x{caps['minCY']} FPS: {unit/caps['maxInterval']:.3f}-{unit/caps['minInterval']:.3f} Format: {format}")
+                else:
+                    print(f"    {caps['id']}: Resolution: {caps['minCX']}x{caps['minCY']}-{caps['maxCX']}x{caps['maxCY']} FPS: {unit/caps['maxInterval']:.3f}-{unit/caps['minInterval']:.3f} Format: {format}")
+    else:
         if args.list_cameras == 1:
-            print(f"{i}: {camera_name}")
+            print("Available cameras:")
-        else:
+        for cam in info:
-            print(camera_name)
+            type = ""
+            if cam['type'] == "Blackmagic":
+                type = "Blackmagic: "
+            if args.list_cameras == 1:
+                print(f"{cam['index']}: {type}{cam['name']}")
+            else:
+                print(f"{type}{cam['name']}")
     cap.destroy_capture()
     sys.exit(0)
 
@@ -61,8 +129,25 @@ import time
 import cv2
 import socket
 import struct
-from input_reader import InputReader, VideoReader
+import json
-from tracker import Tracker
+from input_reader import InputReader, VideoReader, DShowCaptureReader, try_int
+from tracker import Tracker, get_model_base_path
+
+if args.benchmark > 0:
+    model_base_path = get_model_base_path(args.model_dir)
+    im = cv2.imread(os.path.join(model_base_path, "benchmark.bin"), cv2.IMREAD_COLOR)
+    results = []
+    for model_type in [3, 2, 1, 0, -1, -2, -3]:
+        tracker = Tracker(224, 224, threshold=0.1, max_threads=args.max_threads, max_faces=1, discard_after=0, scan_every=0, silent=True, model_type=model_type, model_dir=args.model_dir, no_gaze=(model_type == -1), detection_threshold=0.1, use_retinaface=0, max_feature_updates=900, static_model=True if args.no_3d_adapt == 1 else False)
+        tracker.detected = 1
+        tracker.faces = [(0, 0, 224, 224)]
+        total = 0.0
+        for i in range(100):
+            start = time.perf_counter()
+            r = tracker.predict(im)
+            total += time.perf_counter() - start
+        print(1. / (total / 100.))
+    sys.exit(0)
 
 target_ip = args.ip
 target_port = args.port
@@ -70,13 +155,19 @@ target_port = args.port
 if args.faces >= 40:
     print("Transmission of tracking data over network is not supported with 40 or more faces.")
 
-fps = 0
+fps = args.fps
+dcap = None
+use_dshowcapture_flag = False
 if os.name == 'nt':
-    fps = args.fps
+    dcap = args.dcap
     use_dshowcapture_flag = True if args.use_dshowcapture == 1 else False
-    input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps, use_dshowcapture=use_dshowcapture_flag)
+    input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps, use_dshowcapture=use_dshowcapture_flag, dcap=dcap)
+    if args.dcap == -1 and type(input_reader) == DShowCaptureReader:
+        fps = min(fps, input_reader.device.get_fps())
+elif sys.platform == 'linux' and args.dformat:
+    input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps, dcap=args.dformat)
 else:
-    input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps, use_dshowcapture=False)
+    input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps)
 if type(input_reader.reader) == VideoReader:
     fps = 0
 
@@ -87,6 +178,7 @@ height = 0
 width = 0
 tracker = None
 sock = None
+total_tracking_time = 0.0
 tracking_time = 0.0
 tracking_frames = 0
 frame_count = 0
@@ -96,40 +188,61 @@ features = ["eye_l", "eye_r", "eyebrow_steepness_l", "eyebrow_updown_l", "eyebro
 if args.log_data != "":
     log = open(args.log_data, "w")
     log.write("Frame,Time,Width,Height,FPS,Face,FaceID,RightOpen,LeftOpen,AverageConfidence,Success3D,PnPError,RotationQuat.X,RotationQuat.Y,RotationQuat.Z,RotationQuat.W,Euler.X,Euler.Y,Euler.Z,RVec.X,RVec.Y,RVec.Z,TVec.X,TVec.Y,TVec.Z")
-    for i in range(66):
+    for i in range(68):
         log.write(f",Landmark[{i}].X,Landmark[{i}].Y,Landmark[{i}].Confidence")
-    for i in range(66):
+    for i in range(70):
         log.write(f",Point3D[{i}].X,Point3D[{i}].Y,Point3D[{i}].Z")
     for feature in features:
         log.write(f",{feature}")
     log.write("\r\n")
     log.flush()
 
+is_camera = args.capture == str(try_int(args.capture))
+
 try:
+    attempt = 0
     frame_time = time.perf_counter()
     target_duration = 0
     if fps > 0:
         target_duration = 1. / float(fps)
     repeat = args.repeat_video != 0 and type(input_reader.reader) == VideoReader
     need_reinit = 0
+    failures = 0
+    source_name = input_reader.name
     while repeat or input_reader.is_open():
         if not input_reader.is_open() or need_reinit == 1:
-            input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps)
+            input_reader = InputReader(args.capture, args.raw_rgb, args.width, args.height, fps, use_dshowcapture=use_dshowcapture_flag, dcap=dcap)
+            if input_reader.name != source_name:
+                print(f"Failed to reinitialize camera and got {input_reader.name} instead of {source_name}.")
+                sys.exit(1)
             need_reinit = 2
-            time.sleep(0.001)
+            time.sleep(0.02)
             continue
         if not input_reader.is_ready():
-            time.sleep(0.001)
+            time.sleep(0.02)
             continue
 
         ret, frame = input_reader.read()
+        if ret and args.mirror_input:
+            frame = cv2.flip(frame, 1)
         if not ret:
             if repeat:
                 if need_reinit == 0:
                     need_reinit = 1
                 continue
-            break
+            elif is_camera:
+                attempt += 1
+                if attempt > 30:
+                    break
+                else:
+                    time.sleep(0.02)
+                    if attempt == 3:
+                        need_reinit = 1
+                    continue
+            else:
+                break;
 
+        attempt = 0
         need_reinit = 0
         frame_count += 1
         now = time.time()
@@ -138,179 +251,197 @@ try:
             first = False
             height, width, channels = frame.shape
             sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
-            tracker = Tracker(width, height, threshold=args.threshold, max_threads=args.max_threads, max_faces=args.faces, discard_after=args.discard_after, scan_every=args.scan_every, silent=False if args.silent == 0 else True, model_type=args.model, model_dir=args.model_dir, no_gaze=False if args.gaze_tracking != 0 else True, use_retinaface=args.scan_retinaface, max_feature_updates=args.max_feature_updates, static_model=True if args.no_3d_adapt == 1 else False)
+            tracker = Tracker(width, height, threshold=args.threshold, max_threads=args.max_threads, max_faces=args.faces, discard_after=args.discard_after, scan_every=args.scan_every, silent=False if args.silent == 0 else True, model_type=args.model, model_dir=args.model_dir, no_gaze=False if args.gaze_tracking != 0 and args.model != -1 else True, detection_threshold=args.detection_threshold, use_retinaface=args.scan_retinaface, max_feature_updates=args.max_feature_updates, static_model=True if args.no_3d_adapt == 1 else False, try_hard=args.try_hard == 1)
-            if not args.video_out is None:
+            if args.video_out is not None:
-                out = cv2.VideoWriter(args.video_out, cv2.VideoWriter_fourcc('F','F','V','1'), 24, (width,height))
+                out = cv2.VideoWriter(args.video_out, cv2.VideoWriter_fourcc('F','F','V','1'), args.video_fps, (width * args.video_scale, height * args.video_scale))
 
-        inference_start = time.perf_counter()
+        try:
-        faces = tracker.predict(frame)
+            inference_start = time.perf_counter()
-        if len(faces) > 0:
+            faces = tracker.predict(frame)
-            tracking_time += (time.perf_counter() - inference_start) / len(faces)
+            if len(faces) > 0:
-            tracking_frames += 1
+                inference_time = (time.perf_counter() - inference_start)
-        packet = bytearray()
+                total_tracking_time += inference_time
-        detected = False
+                tracking_time += inference_time / len(faces)
-        for face_num, f in enumerate(faces):
+                tracking_frames += 1
-            f = copy.copy(f)
+            packet = bytearray()
-            f.id += args.face_id_offset
+            detected = False
-            if f.eye_blink is None:
+            for face_num, f in enumerate(faces):
-                f.eye_blink = [1, 1]
+                f = copy.copy(f)
-            right_state = "O" if f.eye_blink[0] > 0.30 else "-"
+                f.id += args.face_id_offset
-            left_state = "O" if f.eye_blink[1] > 0.30 else "-"
+                if f.eye_blink is None:
-            if args.silent == 0:
+                    f.eye_blink = [1, 1]
-                print(f"Confidence[{f.id}]: {f.conf:.4f} / 3D fitting error: {f.pnp_error:.4f} / Eyes: {left_state}, {right_state}")
+                right_state = "O" if f.eye_blink[0] > 0.30 else "-"
-            detected = True
+                left_state = "O" if f.eye_blink[1] > 0.30 else "-"
-            if not f.success:
+                if args.silent == 0:
-                pts_3d = np.zeros((70, 3), np.float32)
+                    print(f"Confidence[{f.id}]: {f.conf:.4f} / 3D fitting error: {f.pnp_error:.4f} / Eyes: {left_state}, {right_state}")
-            packet.extend(bytearray(struct.pack("d", now)))
+                detected = True
-            packet.extend(bytearray(struct.pack("i", f.id)))
+                if not f.success:
-            packet.extend(bytearray(struct.pack("f", width)))
+                    pts_3d = np.zeros((70, 3), np.float32)
-            packet.extend(bytearray(struct.pack("f", height)))
+                packet.extend(bytearray(struct.pack("d", now)))
-            packet.extend(bytearray(struct.pack("f", f.eye_blink[0])))
+                packet.extend(bytearray(struct.pack("i", f.id)))
-            packet.extend(bytearray(struct.pack("f", f.eye_blink[1])))
+                packet.extend(bytearray(struct.pack("f", width)))
-            packet.extend(bytearray(struct.pack("B", 1 if f.success else 0)))
+                packet.extend(bytearray(struct.pack("f", height)))
-            packet.extend(bytearray(struct.pack("f", f.pnp_error)))
+                packet.extend(bytearray(struct.pack("f", f.eye_blink[0])))
-            packet.extend(bytearray(struct.pack("f", f.quaternion[0])))
+                packet.extend(bytearray(struct.pack("f", f.eye_blink[1])))
-            packet.extend(bytearray(struct.pack("f", f.quaternion[1])))
+                packet.extend(bytearray(struct.pack("B", 1 if f.success else 0)))
-            packet.extend(bytearray(struct.pack("f", f.quaternion[2])))
+                packet.extend(bytearray(struct.pack("f", f.pnp_error)))
-            packet.extend(bytearray(struct.pack("f", f.quaternion[3])))
+                packet.extend(bytearray(struct.pack("f", f.quaternion[0])))
-            packet.extend(bytearray(struct.pack("f", f.euler[0])))
+                packet.extend(bytearray(struct.pack("f", f.quaternion[1])))
-            packet.extend(bytearray(struct.pack("f", f.euler[1])))
+                packet.extend(bytearray(struct.pack("f", f.quaternion[2])))
-            packet.extend(bytearray(struct.pack("f", f.euler[2])))
+                packet.extend(bytearray(struct.pack("f", f.quaternion[3])))
-            packet.extend(bytearray(struct.pack("f", f.translation[0])))
+                packet.extend(bytearray(struct.pack("f", f.euler[0])))
-            packet.extend(bytearray(struct.pack("f", f.translation[1])))
+                packet.extend(bytearray(struct.pack("f", f.euler[1])))
-            packet.extend(bytearray(struct.pack("f", f.translation[2])))
+                packet.extend(bytearray(struct.pack("f", f.euler[2])))
-            if not log is None:
+                packet.extend(bytearray(struct.pack("f", f.translation[0])))
-                log.write(f"{frame_count},{now},{width},{height},{args.fps},{face_num},{f.id},{f.eye_blink[0]},{f.eye_blink[1]},{f.conf},{f.success},{f.pnp_error},{f.quaternion[0]},{f.quaternion[1]},{f.quaternion[2]},{f.quaternion[3]},{f.euler[0]},{f.euler[1]},{f.euler[2]},{f.rotation[0]},{f.rotation[1]},{f.rotation[2]},{f.translation[0]},{f.translation[1]},{f.translation[2]}")
+                packet.extend(bytearray(struct.pack("f", f.translation[1])))
-            for (x,y,c) in f.lms:
+                packet.extend(bytearray(struct.pack("f", f.translation[2])))
-                packet.extend(bytearray(struct.pack("f", c)))
+                if log is not None:
-            if args.visualize > 1:
+                    log.write(f"{frame_count},{now},{width},{height},{fps},{face_num},{f.id},{f.eye_blink[0]},{f.eye_blink[1]},{f.conf},{f.success},{f.pnp_error},{f.quaternion[0]},{f.quaternion[1]},{f.quaternion[2]},{f.quaternion[3]},{f.euler[0]},{f.euler[1]},{f.euler[2]},{f.rotation[0]},{f.rotation[1]},{f.rotation[2]},{f.translation[0]},{f.translation[1]},{f.translation[2]}")
-                frame = cv2.putText(frame, str(f.id), (int(f.bbox[0]), int(f.bbox[1])), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (255,0,255))
+                for (x,y,c) in f.lms:
-            if args.visualize > 2:
+                    packet.extend(bytearray(struct.pack("f", c)))
-                frame = cv2.putText(frame, f"{f.conf:.4f}", (int(f.bbox[0] + 18), int(f.bbox[1] - 6)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0))
+                if args.visualize > 1:
-            for pt_num, (x,y,c) in enumerate(f.lms):
+                    frame = cv2.putText(frame, str(f.id), (int(f.bbox[0]), int(f.bbox[1])), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (255,0,255))
-                packet.extend(bytearray(struct.pack("f", y)))
+                if args.visualize > 2:
-                packet.extend(bytearray(struct.pack("f", x)))
+                    frame = cv2.putText(frame, f"{f.conf:.4f}", (int(f.bbox[0] + 18), int(f.bbox[1] - 6)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,0,255))
-                if not log is None:
+                for pt_num, (x,y,c) in enumerate(f.lms):
-                    log.write(f",{y},{x},{c}")
+                    packet.extend(bytearray(struct.pack("f", y)))
-                if pt_num == 66 and f.eye_blink[0] < 0.30:
+                    packet.extend(bytearray(struct.pack("f", x)))
-                    continue
+                    if log is not None:
-                if pt_num == 67 and f.eye_blink[1] < 0.30:
+                        log.write(f",{y},{x},{c}")
-                    continue
+                    if pt_num == 66 and (f.eye_blink[0] < 0.30 or c < 0.20):
-                x = int(x + 0.5)
+                        continue
-                y = int(y + 0.5)
+                    if pt_num == 67 and (f.eye_blink[1] < 0.30 or c < 0.20):
-                if args.visualize != 0 or not out is None:
+                        continue
-                    if args.visualize > 3:
-                        frame = cv2.putText(frame, str(pt_num), (int(y), int(x)), cv2.FONT_HERSHEY_SIMPLEX, 0.25, (255,255,0))
-                    color = (0, 0, 255)
-                    if pt_num >= 66:
-                        color = (255, 255, 0)
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
-                        frame[int(x), int(y)] = color
-                    x += 1
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
-                        frame[int(x), int(y)] = color
-                    y += 1
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
-                        frame[int(x), int(y)] = color
-                    x -= 1
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
-                        frame[int(x), int(y)] = color
-            if args.pnp_points != 0 and (args.visualize != 0 or not out is None) and f.rotation is not None:
-                if args.pnp_points > 1:
-                    projected = cv2.projectPoints(f.face_3d[0:66], f.rotation, f.translation, tracker.camera, tracker.dist_coeffs)
-                else:
-                    projected = cv2.projectPoints(f.contour, f.rotation, f.translation, tracker.camera, tracker.dist_coeffs)
-                for [(x,y)] in projected[0]:
                     x = int(x + 0.5)
                     y = int(y + 0.5)
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
+                    if args.visualize != 0 or out is not None:
-                        frame[int(x), int(y)] = (0, 255, 255)
+                        if args.visualize > 3:
-                    x += 1
+                            frame = cv2.putText(frame, str(pt_num), (int(y), int(x)), cv2.FONT_HERSHEY_SIMPLEX, 0.25, (255,255,0))
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
+                        color = (0, 255, 0)
-                        frame[int(x), int(y)] = (0, 255, 255)
+                        if pt_num >= 66:
-                    y += 1
+                            color = (255, 255, 0)
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
+                        if not (x < 0 or y < 0 or x >= height or y >= width):
-                        frame[int(x), int(y)] = (0, 255, 255)
+                            cv2.circle(frame, (y, x), 1, color, -1)
-                    x -= 1
+                if args.pnp_points != 0 and (args.visualize != 0 or out is not None) and f.rotation is not None:
-                    if not (x < 0 or y < 0 or x >= height or y >= width):
+                    if args.pnp_points > 1:
-                        frame[int(x), int(y)] = (0, 255, 255)
+                        projected = cv2.projectPoints(f.face_3d[0:66], f.rotation, f.translation, tracker.camera, tracker.dist_coeffs)
-            for (x,y,z) in f.pts_3d:
+                    else:
-                packet.extend(bytearray(struct.pack("f", x)))
+                        projected = cv2.projectPoints(f.contour, f.rotation, f.translation, tracker.camera, tracker.dist_coeffs)
-                packet.extend(bytearray(struct.pack("f", -y)))
+                    for [(x,y)] in projected[0]:
-                packet.extend(bytearray(struct.pack("f", -z)))
+                        x = int(x + 0.5)
-                if not log is None:
+                        y = int(y + 0.5)
-                    log.write(f",{x},{-y},{-z}")
+                        if not (x < 0 or y < 0 or x >= height or y >= width):
-            if f.current_features is None:
+                            frame[int(x), int(y)] = (0, 255, 255)
-                f.current_features = {}
+                        x += 1
-            for feature in features:
+                        if not (x < 0 or y < 0 or x >= height or y >= width):
-                if not feature in f.current_features:
+                            frame[int(x), int(y)] = (0, 255, 255)
-                    f.current_features[feature] = 0
+                        y += 1
-                packet.extend(bytearray(struct.pack("f", f.current_features[feature])))
+                        if not (x < 0 or y < 0 or x >= height or y >= width):
-                if not log is None:
+                            frame[int(x), int(y)] = (0, 255, 255)
-                    log.write(f",{f.current_features[feature]}")
+                        x -= 1
-            if not log is None:
+                        if not (x < 0 or y < 0 or x >= height or y >= width):
-                log.write("\r\n")
+                            frame[int(x), int(y)] = (0, 255, 255)
-                log.flush()
+                for (x,y,z) in f.pts_3d:
+                    packet.extend(bytearray(struct.pack("f", x)))
+                    packet.extend(bytearray(struct.pack("f", -y)))
+                    packet.extend(bytearray(struct.pack("f", -z)))
+                    if log is not None:
+                        log.write(f",{x},{-y},{-z}")
+                if f.current_features is None:
+                    f.current_features = {}
+                for feature in features:
+                    if not feature in f.current_features:
+                        f.current_features[feature] = 0
+                    packet.extend(bytearray(struct.pack("f", f.current_features[feature])))
+                    if log is not None:
+                        log.write(f",{f.current_features[feature]}")
+                if log is not None:
+                    log.write("\r\n")
+                    log.flush()
 
-        if detected and len(faces) < 40:
+            if detected and len(faces) < 40:
-            sock.sendto(packet, (target_ip, target_port))
+                sock.sendto(packet, (target_ip, target_port))
 
-        if not out is None:
+            if out is not None:
-            out.write(frame)
+                video_frame = frame
+                if args.video_scale != 1:
+                    video_frame = cv2.resize(frame, (width * args.video_scale, height * args.video_scale), interpolation=cv2.INTER_NEAREST)
+                out.write(video_frame)
+                if args.video_scale != 1:
+                    del video_frame
 
-        if args.visualize != 0:
+            if args.visualize != 0:
-            cv2.imshow('OpenSeeFace Visualization', frame)
+                cv2.imshow('OpenSeeFace Visualization', frame)
-            if cv2.waitKey(1) & 0xFF == ord('q'):
+                if cv2.waitKey(1) & 0xFF == ord('q'):
-                if args.dump_points != "" and not faces is None and len(faces) > 0:
+                    if args.dump_points != "" and faces is not None and len(faces) > 0:
-                    np.set_printoptions(threshold=sys.maxsize, precision=15)
+                        np.set_printoptions(threshold=sys.maxsize, precision=15)
-                    pairs = [
+                        pairs = [
-                        (0, 16),
+                            (0, 16),
-                        (1, 15),
+                            (1, 15),
-                        (2, 14),
+                            (2, 14),
-                        (3, 13),
+                            (3, 13),
-                        (4, 12),
+                            (4, 12),
-                        (5, 11),
+                            (5, 11),
-                        (6, 10),
+                            (6, 10),
-                        (7, 9),
+                            (7, 9),
-                        (17, 26),
+                            (17, 26),
-                        (18, 25),
+                            (18, 25),
-                        (19, 24),
+                            (19, 24),
-                        (20, 23),
+                            (20, 23),
-                        (21, 22),
+                            (21, 22),
-                        (31, 35),
+                            (31, 35),
-                        (32, 34),
+                            (32, 34),
-                        (36, 45),
+                            (36, 45),
-                        (37, 44),
+                            (37, 44),
-                        (38, 43),
+                            (38, 43),
-                        (39, 42),
+                            (39, 42),
-                        (40, 47),
+                            (40, 47),
-                        (41, 46),
+                            (41, 46),
-                        (48, 52),
+                            (48, 52),
-                        (49, 51),
+                            (49, 51),
-                        (56, 54),
+                            (56, 54),
-                        (57, 53),
+                            (57, 53),
-                        (58, 62),
+                            (58, 62),
-                        (59, 61),
+                            (59, 61),
-                        (65, 63)
+                            (65, 63)
-                    ]
+                        ]
-                    points = copy.copy(faces[0].face_3d)
+                        points = copy.copy(faces[0].face_3d)
-                    for a, b in pairs:
+                        for a, b in pairs:
-                        x = (points[a, 0] - points[b, 0]) / 2.0
+                            x = (points[a, 0] - points[b, 0]) / 2.0
-                        y = (points[a, 1] + points[b, 1]) / 2.0
+                            y = (points[a, 1] + points[b, 1]) / 2.0
-                        z = (points[a, 2] + points[b, 2]) / 2.0
+                            z = (points[a, 2] + points[b, 2]) / 2.0
-                        points[a, 0] = x
+                            points[a, 0] = x
-                        points[b, 0] = -x
+                            points[b, 0] = -x
-                        points[[a, b], 1] = y
+                            points[[a, b], 1] = y
-                        points[[a, b], 2] = z
+                            points[[a, b], 2] = z
-                    points[[8, 27, 28, 29, 33, 50, 55, 60, 64], 0] = 0.0
+                        points[[8, 27, 28, 29, 33, 50, 55, 60, 64], 0] = 0.0
-                    points[30, :] = 0.0
+                        points[30, :] = 0.0
-                    with open(args.dump_points, "w") as fh:
+                        with open(args.dump_points, "w") as fh:
-                        fh.write(repr(points))
+                            fh.write(repr(points))
+                    break
+            failures = 0
+        except Exception as e:
+            if e.__class__ == KeyboardInterrupt:
+                if args.silent == 0:
+                    print("Quitting")
                 break
+            traceback.print_exc()
+            failures += 1
+            if failures > 30:
+                break
+
+        collected = False
+        del frame
 
         duration = time.perf_counter() - frame_time
         while duration < target_duration:
-            time.sleep(target_duration - duration)
+            if not collected:
+                gc.collect()
+                collected = True
+            duration = time.perf_counter() - frame_time
+            sleep_time = target_duration - duration
+            if sleep_time > 0:
+                time.sleep(sleep_time)
             duration = time.perf_counter() - frame_time
         frame_time = time.perf_counter()
 except KeyboardInterrupt:
@@ -318,10 +449,12 @@ except KeyboardInterrupt:
         print("Quitting")
 
 input_reader.close()
-if not out is None:
+if out is not None:
     out.release()
-cv2.destroyAllWindows()
+if args.visualize != 0:
+    cv2.destroyAllWindows()
 
 if args.silent == 0 and tracking_frames > 0:
-    tracking_time = 1000 * tracking_time / tracking_frames
+    average_tracking_time = 1000 * tracking_time / tracking_frames
-    print(f"Average tracking time per detected face: {tracking_time:.2f} ms")
+    print(f"Average tracking time per detected face: {average_tracking_time:.2f} ms")
+    print(f"Tracking time: {total_tracking_time:.3f} s\nFrames: {tracking_frames}")

facetracker.spec

@@ -0,0 +1,53 @@
+# -*- mode: python ; coding: utf-8 -*-
+
+block_cipher = None
+
+
+a = Analysis(['facetracker.py'],
+             pathex=['C:\\OpenSeeFaceBuild'],
+             binaries=[('dshowcapture/dshowcapture_x86.dll', '.'), ('dshowcapture/dshowcapture_x64.dll', '.'), ('dshowcapture/libminibmcapture32.dll', '.'), ('dshowcapture/libminibmcapture64.dll', '.'), ('escapi/escapi_x86.dll', '.'), ('escapi/escapi_x64.dll', '.'), ('run.bat', '.'), ('msvcp140.dll', '.'), ('vcomp140.dll', '.'), ('concrt140.dll', '.'), ('vccorlib140.dll', '.')],
+             datas=[],
+             hiddenimports=[],
+             hookspath=[],
+             runtime_hooks=[],
+             excludes=['matplotlib', 'mpl-data', 'PyInstaller', 'pywt', 'skimage', 'scipy', 'pyinstaller'],
+             win_no_prefer_redirects=False,
+             win_private_assemblies=False,
+             cipher=block_cipher,
+             noarchive=False)
+
+remove_bin = []
+for bin in a.binaries:
+    if bin[0].startswith("opencv_video") or bin[0].startswith("PyInstaller"):
+        remove_bin.append(bin[0])
+        print(bin)
+remove_dat = []
+for bin in a.datas:
+    if bin[0].startswith("opencv_video") or bin[0].startswith("PyInstaller"):
+        remove_dat.append(bin[0])
+        print("data ", bin)
+
+a.binaries = [x for x in a.binaries if not x[0] in remove_bin]
+a.datas = [x for x in a.datas if not x[0] in remove_dat]
+
+pyz = PYZ(a.pure, a.zipped_data,
+             cipher=block_cipher)
+exe = EXE(pyz,
+          a.scripts,
+          [],
+          exclude_binaries=True,
+          name='facetracker',
+          debug=False,
+          bootloader_ignore_signals=False,
+          strip=False,
+          upx=True,
+          console=True )
+
+coll = COLLECT(exe,
+               a.binaries,
+               a.zipfiles,
+               a.datas,
+               strip=False,
+               upx=True,
+               upx_exclude=[],
+               name='facetracker')

input_reader.py

@@ -7,6 +7,7 @@ import escapi
 import dshowcapture
 import time
 import traceback
+import gc
 
 class VideoReader():
     def __init__(self, capture, camera=False):
@@ -17,6 +18,7 @@ class VideoReader():
         if self.cap is None:
             print("The video source cannot be opened")
             sys.exit(0)
+        self.name = str(capture)
     def is_open(self):
         return self.cap.isOpened()
     def is_ready(self):
@@ -33,6 +35,8 @@ class EscapiReader(VideoReader):
         self.height = height
         self.fps = fps
         self.device = capture
+        escapi.count_capture_devices()
+        self.name = str(escapi.device_name(self.device).decode('utf8', 'surrogateescape'))
         self.buffer = escapi.init_camera(self.device, self.width, self.height, self.fps)
         escapi.do_capture(self.device)
     def is_open(self):
@@ -50,22 +54,46 @@ class EscapiReader(VideoReader):
         escapi.deinit_camera(self.device)
 
 class DShowCaptureReader(VideoReader):
-    def __init__(self, capture, width, height, fps, use_dshowcapture=True):
+    def __init__(self, capture, width, height, fps, use_dshowcapture=True, dcap=None):
         self.device = None
         self.width = width
         self.height = height
         self.fps = fps
+        self.dcap = dcap;
         self.device = dshowcapture.DShowCapture()
         self.device.get_devices()
-        self.device_name = self.device.get_device(capture)
+        info = self.device.get_info()
-        self.device.capture_device(capture, self.width, self.height, self.fps)
+        self.name = info[capture]['name']
+        if info[capture]['type'] == "Blackmagic":
+            self.name = "Blackmagic: " + self.name
+            if dcap is None or dcap < 0:
+                dcap = 0
+        ret = False
+        if dcap is None:
+            ret = self.device.capture_device(capture, self.width, self.height, self.fps)
+        else:
+            if dcap < 0:
+                ret = self.device.capture_device_default(capture)
+            else:
+                ret = self.device.capture_device_by_dcap(capture, dcap, self.width, self.height, self.fps)
+        if not ret:
+            raise Exception("Failed to start capture.")
+        self.width = self.device.width
+        self.height = self.device.height
+        self.fps = self.device.fps
+        print(f"Camera: \"{self.name}\" Capability ID: {dcap} Resolution: {self.device.width}x{self.device.height} Frame rate: {self.device.fps} Colorspace: {self.device.colorspace} Internal: {self.device.colorspace_internal} Flipped: {self.device.flipped}")
         self.timeout = 1000
     def is_open(self):
         return self.device.capturing()
     def is_ready(self):
         return self.device.capturing()
     def read(self):
-        img = self.device.get_frame(self.timeout)
+        img = None
+        try:
+            img = self.device.get_frame(self.timeout)
+        except:
+            gc.collect()
+            img = self.device.get_frame(self.timeout)
         if img is None:
             return False, None
         else:
@@ -79,9 +107,11 @@ class OpenCVReader(VideoReader):
         self.width = width
         self.height = height
         self.fps = fps
+        self.name = str(capture)
         super(OpenCVReader, self).__init__(capture, camera=True)
         self.cap.set(3, width)
         self.cap.set(4, height)
+        self.cap.set(38, 1)
     def is_open(self):
         return super(OpenCVReader, self).is_open()
     def is_ready(self):
@@ -91,6 +121,30 @@ class OpenCVReader(VideoReader):
     def close(self):
         super(OpenCVReader, self).close()
 
+class V4L2Reader(OpenCVReader):
+    def __init__(self, capture, width, height, fps, dformat):
+        self.cap = cv2.VideoCapture(capture, cv2.CAP_V4L2)
+        if dformat:
+            fourcc = cv2.VideoWriter.fourcc(*dformat)
+            self.cap.set(6, fourcc)
+        self.device = None
+        self.width = width
+        self.height = height
+        self.fps = fps
+        self.name = str(capture)
+        self.cap.set(3, width)
+        self.cap.set(4, height)
+        self.cap.set(38, 1)
+        self.cap.set(5, fps)
+    def is_open(self):
+        return super(V4L2Reader, self).is_open()
+    def is_ready(self):
+        return super(V4L2Reader, self).is_ready()
+    def read(self):
+        return super(V4L2Reader, self).read()
+    def close(self):
+        super(V4L2Reader, self).close()
+
 class RawReader:
     def __init__(self, width, height):
         self.width = int(width)
@@ -124,7 +178,7 @@ def try_int(s):
         return None
 
 def test_reader(reader):
-    got_any = False
+    got_any = 0
     try:
         for i in range(30):
             if not reader.is_ready():
@@ -135,9 +189,11 @@ def test_reader(reader):
                 print("No frame")
             else:
                 print("Got frame")
-                got_any = True
+                got_any += 1
+                if got_any > 10:
+                    break
         if reader.is_open():
-            return got_any
+            return got_any > 0
         print("Fail")
         return False
     except:
@@ -146,8 +202,9 @@ def test_reader(reader):
         return False
 
 class InputReader():
-    def __init__(self, capture, raw_rgb, width, height, fps, use_dshowcapture=False):
+    def __init__(self, capture, raw_rgb, width, height, fps, use_dshowcapture=False, dcap=None):
         self.reader = None
+        self.name = str(capture)
         try:
             if raw_rgb > 0:
                 self.reader = RawReader(width, height)
@@ -160,12 +217,15 @@ class InputReader():
                     name = ""
                     try:
                         if use_dshowcapture:
-                            self.reader = DShowCaptureReader(int(capture), width, height, fps)
+                            self.reader = DShowCaptureReader(int(capture), width, height, fps, dcap=dcap)
-                            name = self.reader.device_name
+                            name = self.reader.name
                             good = test_reader(self.reader)
+                            self.name = name
                         else:
                             good = False
                     except:
+                        print("DShowCapture exception: ")
+                        traceback.print_exc()
                         good = False
                     if good:
                         return
@@ -177,7 +237,7 @@ class InputReader():
                         devices = escapi.count_capture_devices()
                         found = None
                         for i in range(devices):
-                            escapi_name = str(escapi.device_name(i).decode('utf8'))
+                            escapi_name = str(escapi.device_name(i).decode('utf8', 'surrogateescape'))
                             if name == escapi_name:
                                 found = i
                         if found is None:
@@ -187,12 +247,23 @@ class InputReader():
                             self.reader = EscapiReader(found, width, height, fps)
                             good = test_reader(self.reader)
                     except:
+                        print("Escapi exception: ")
+                        traceback.print_exc()
                         good = False
                     if good:
                         return
                     # Try with OpenCV
                     print(f"Escapi failed. Falling back to OpenCV. If this fails, please change your camera settings.", file=sys.stderr)
                     self.reader = OpenCVReader(int(capture), width, height, fps)
+                    self.name = self.reader.name
+                elif sys.platform == 'linux':
+                    try:
+                        self.reader = V4L2Reader(int(capture), width, height, fps, dcap)
+                    except:
+                        print("V4L2 exception: ")
+                        traceback.print_exc()
+                        print(f"V4L2 failed. Falling back to OpenCV. If this fails, please change your camera settings.", file=sys.stderr)
+                        self.reader = OpenCVReader(int(capture), width, height, fps)
                 else:
                     self.reader = OpenCVReader(int(capture), width, height, fps)
         except Exception as e:

make_exe.bat

@@ -1,13 +1,31 @@
-REM pyinstaller --onefile ^
+echo "Started"
-REM     --add-binary dshowcapture/dshowcapture_x86.dll;. ^
-REM     --add-binary dshowcapture/dshowcapture_x64.dll;. ^
-REM     --add-binary escapi/escapi_x86.dll;. ^
-REM     --add-binary escapi/escapi_x64.dll;. ^
-REM     facetracker.py
 
-pyinstaller --onedir ^
+:: A venv is used so it's easy to add the onnxruntime dll to the binary
-    --add-binary dshowcapture/dshowcapture_x86.dll;. ^
+
-    --add-binary dshowcapture/dshowcapture_x64.dll;. ^
+echo "Creating venv"
-    --add-binary escapi/escapi_x86.dll;. ^
+python -m venv venv
-    --add-binary escapi/escapi_x64.dll;. ^
+
-    facetracker.py
+echo "Activating venv"
+call venv\Scripts\activate.bat
+
+echo "Installing dependencies"
+pip install wheel
+
+pip install onnxruntime opencv-python==4.5.4.60 pillow numpy==1.23.0 pyinstaller
+
+echo "Running pyinstaller"
+pyinstaller facetracker.py --clean ^
+    --onedir ^
+    --add-binary dshowcapture/*.dll;. ^
+    --add-binary escapi/*.dll;. ^
+    --add-binary venv/lib/site-packages/onnxruntime/capi/*.dll;onnxruntime\capi ^
+    --add-binary msvcp140.dll;. ^
+    --add-binary vcomp140.dll;. ^
+    --add-binary concrt140.dll;. ^
+    --add-binary vccorlib140.dll;. ^
+    --add-binary run.bat;.
+
+echo "Deleting opencv dll"
+del dist\facetracker\cv2\opencv_videoio_*
+
+echo "Finished"

make_exe.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+set -e
+
+PYTHON_BINARY=$(which python3)
+
+if [[ -z "$PYTHON_BINARY" ]]; then
+    echo "python3 not found. Searching for python instead."
+    PYTHON_BINARY=$(which python)
+    if [[ -z "$PYTHON_BINARY" ]]; then
+        echo "Python 3 must be available on your path. Exiting."
+        exit 1
+    fi
+fi
+
+echo "Creating venv"
+"$PYTHON_BINARY" -m venv venv
+
+source venv/bin/activate
+
+echo "Installing packages"
+pip install wheel # Make sure this is installed beforehand
+
+pip install onnxruntime opencv-python==4.5.4.60 pillow numpy==1.23.0 pyinstaller
+
+echo "Creating binary"
+pyinstaller --onedir --clean facetracker.py \
+    --add-binary venv/lib/python3.*/site-packages/onnxruntime/capi/*.so:onnxruntime/capi
+
+echo "Files should be available in the dist/ folder"
+echo "Done!"
+

model.py

@@ -1,7 +1,7 @@
 # This file is not used by the tracking application and currently outdated
 import torch
 import torch.nn as nn
-import geffnet.mobilenetv3 # geffnet.mobilenetv3._gen_mobilenet_v3 needs to be patched to return the parameters instead of instantiating the network
+import geffnet.mobilenetv3 # geffnet.mobilenetv3._gen_mobilenet_v3 needs to be patched to return the parameters instead of instantiating the network, based on commit c450c12ae6ffb1757f62dde3c2765da3c10f6def
 from geffnet.efficientnet_builder import round_channels
 
 class DSConv2d(nn.Module):
@@ -121,6 +121,10 @@ class OpenSeeFaceDetect(geffnet.mobilenetv3.MobileNetV3):
     def forward(self, x):
         return self._forward_impl(x)
 
+def logit_arr(p, factor=16.0):
+    p = p.clamp(0.0000001, 0.9999999)
+    return torch.log(p / (1 - p)) / factor
+
 # Landmark detection model
 # Models:
 # 0: "small", 0.5
@@ -128,7 +132,7 @@ class OpenSeeFaceDetect(geffnet.mobilenetv3.MobileNetV3):
 # 2: "large", 0.75
 # 3: "large", 1.0
 class OpenSeeFaceLandmarks(geffnet.mobilenetv3.MobileNetV3):
-    def __init__(self, size="large", channel_multiplier=1.0):
+    def __init__(self, size="large", channel_multiplier=1.0, inference=False):
         kwargs = geffnet.mobilenetv3._gen_mobilenet_v3([size], channel_multiplier=channel_multiplier)
         super(OpenSeeFaceLandmarks, self).__init__(**kwargs)
         if size == "large":
@@ -143,6 +147,7 @@ class OpenSeeFaceLandmarks(geffnet.mobilenetv3.MobileNetV3):
             self.group = DSConv2d(198 * 1, 198, kernels_per_layer=4, groups=3)
             self.r2_i = 1
             self.r3_i = 2
+        self.inference = inference
     def _forward_impl(self, x):
         x = self.conv_stem(x)
         x = self.bn1(x)
@@ -158,6 +163,62 @@ class OpenSeeFaceLandmarks(geffnet.mobilenetv3.MobileNetV3):
         x = self.up1(x, r3)
         x = self.up2(x, r2)
         x = self.group(x)
+
+        if self.inference:
+            t_main = x[:, 0:66].reshape((-1, 66, 28*28))
+            t_m = t_main.argmax(dim=2)
+            indices = t_m.unsqueeze(2)
+            t_conf = t_main.gather(2, indices).squeeze(2)
+            t_off_x = x[:, 66:132].reshape((-1, 66, 28*28)).gather(2, indices).squeeze(2)
+            t_off_y = x[:, 132:198].reshape((-1, 66, 28*28)).gather(2, indices).squeeze(2)
+            t_off_x = (223. * logit_arr(t_off_x) + 0.5).floor()
+            t_off_y = (223. * logit_arr(t_off_y) + 0.5).floor()
+            t_x = 223. * (t_m / 28.).floor() / 27. + t_off_x
+            t_y = 223. * t_m.remainder(28.).float() / 27. + t_off_y
+            x = (t_conf.mean(1), torch.stack([t_x, t_y, t_conf], 2))
+
+        return x
+    def forward(self, x):
+        return self._forward_impl(x)
+
+# lm_modelT for 56x56 30 point inference
+class OpenSeeFaceLandmarks30Pt(geffnet.mobilenetv3.MobileNetV3):
+    def __init__(self, size="large", channel_multiplier=1.0, inference=False):
+        kwargs = geffnet.mobilenetv3._gen_mobilenet_v3([size], channel_multiplier=channel_multiplier)
+        super(OpenSeeFaceLandmarks30Pt, self).__init__(**kwargs)
+        self.up1 = UNetUp(960, 112, 256, (4,4))
+        self.up2 = UNetUp(256, 40, 180, (7,7))
+        self.group = DSConv2d(180, 90, kernels_per_layer=4, groups=3)
+        self.inference = inference
+    def _forward_impl(self, x):
+        x = self.conv_stem(x)
+        x = self.bn1(x)
+        x = self.act1(x)
+        r2 = None
+        r3 = None
+        for i, feature in enumerate(self.blocks):
+            x = feature(x)
+            if i == 4:
+                r3 = x
+            if i == 2:
+                r2 = x
+        x = self.up1(x, r3)
+        x = self.up2(x, r2)
+        x = self.group(x)
+
+        if self.inference:
+            t_main = x[:, 0:30].reshape((-1, 30, 7*7))
+            t_m = t_main.argmax(dim=2)
+            indices = t_m.unsqueeze(2)
+            t_conf = t_main.gather(2, indices).squeeze(2)
+            t_off_x = x[:, 30:60].reshape((-1, 30, 7*7)).gather(2, indices).squeeze(2)
+            t_off_y = x[:, 60:90].reshape((-1, 30, 7*7)).gather(2, indices).squeeze(2)
+            t_off_x = 55. * logit_arr(t_off_x, factor=8.0)
+            t_off_y = 55. * logit_arr(t_off_y, factor=8.0)
+            t_x = 55. * (t_m / 7.).floor() / 6. + t_off_x
+            t_y = 55. * t_m.remainder(7.).float() / 6. + t_off_y
+            x = (t_conf.mean(1), torch.stack([t_x, t_y, t_conf], 2))
+
         return x
     def forward(self, x):
         return self._forward_impl(x)
@@ -190,7 +251,7 @@ def AdapWingLoss(pre_hm, gt_hm):
         img_dilate = cv2.morphologyEx(img_merge, cv2.MORPH_DILATE, kernel)
         img_dilate[img_dilate < 51] = 1  # 0*W+1
         img_dilate[img_dilate >= 51] = 11  # 1*W+1
-        img_dilate = np.array(img_dilate, dtype=np.int)
+        img_dilate = np.array(img_dilate, dtype=int)
         img_dilate = img_dilate.transpose(2, 0, 1)
         dilated = torch.from_numpy(img_dilate).float()
         if first:
@@ -247,3 +308,7 @@ if __name__== "__main__":
     m=OpenSeeFaceLandmarks("large", 1.0)
     ckpt = torch.load("lm_model3.pth")
     m.load_state_dict(ckpt)
+    print("Checking lm_modelT model")
+    m=OpenSeeFaceLandmarks("large", 1.0)
+    ckpt = torch.load("lm_modelT.pth")
+    m.load_state_dict(ckpt)

poetry.lock

@@ -0,0 +1,224 @@
+[[package]]
+name = "flatbuffers"
+version = "2.0"
+description = "The FlatBuffers serialization format for Python"
+category = "main"
+optional = false
+python-versions = "*"
+
+[[package]]
+name = "numpy"
+version = "1.22"
+description = "NumPy is the fundamental package for array computing with Python."
+category = "main"
+optional = false
+python-versions = ">=3.7,<3.11"
+
+[[package]]
+name = "onnxruntime"
+version = "1.9.0"
+description = "ONNX Runtime is a runtime accelerator for Machine Learning models"
+category = "main"
+optional = false
+python-versions = "*"
+
+[package.dependencies]
+flatbuffers = "*"
+numpy = ">=1.16.6"
+protobuf = "*"
+
+[[package]]
+name = "opencv-python"
+version = "4.5.4.58"
+description = "Wrapper package for OpenCV python bindings."
+category = "main"
+optional = false
+python-versions = ">=3.6"
+
+[package.dependencies]
+numpy = ">=1.21.2"
+
+[[package]]
+name = "pillow"
+version = "10.2.0"
+description = "Python Imaging Library (Fork)"
+category = "main"
+optional = false
+python-versions = ">=3.6"
+
+[[package]]
+name = "protobuf"
+version = "4.25.8"
+description = "Protocol Buffers"
+category = "main"
+optional = false
+python-versions = ">=3.5"
+
+[metadata]
+lock-version = "1.1"
+python-versions = ">=3.7,<3.11"
+content-hash = "1511681290c580145b1117beffe728052d856eee3ad8419930626c6fc6c9a09d"
+
+[metadata.files]
+flatbuffers = [
+    {file = "flatbuffers-2.0-py2.py3-none-any.whl", hash = "sha256:3751954f0604580d3219ae49a85fafec9d85eec599c0b96226e1bc0b48e57474"},
+    {file = "flatbuffers-2.0.tar.gz", hash = "sha256:12158ab0272375eab8db2d663ae97370c33f152b27801fa6024e1d6105fd4dd2"},
+]
+numpy = [
+    {file = "numpy-1.21.3-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:508b0b513fa1266875524ba8a9ecc27b02ad771fe1704a16314dc1a816a68737"},
+    {file = "numpy-1.21.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5dfe9d6a4c39b8b6edd7990091fea4f852888e41919d0e6722fe78dd421db0eb"},
+    {file = "numpy-1.21.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8a10968963640e75cc0193e1847616ab4c718e83b6938ae74dea44953950f6b7"},
+    {file = "numpy-1.21.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:49c6249260890e05b8111ebfc391ed58b3cb4b33e63197b2ec7f776e45330721"},
+    {file = "numpy-1.21.3-cp310-cp310-win_amd64.whl", hash = "sha256:f8f4625536926a155b80ad2bbff44f8cc59e9f2ad14cdda7acf4c135b4dc8ff2"},
+    {file = "numpy-1.21.3-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:e54af82d68ef8255535a6cdb353f55d6b8cf418a83e2be3569243787a4f4866f"},
+    {file = "numpy-1.21.3-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:f41b018f126aac18583956c54544db437f25c7ee4794bcb23eb38bef8e5e192a"},
+    {file = "numpy-1.21.3-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:50cd26b0cf6664cb3b3dd161ba0a09c9c1343db064e7c69f9f8b551f5104d654"},
+    {file = "numpy-1.21.3-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4cc9b512e9fb590797474f58b7f6d1f1b654b3a94f4fa8558b48ca8b3cfc97cf"},
+    {file = "numpy-1.21.3-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:88a5d6b268e9ad18f3533e184744acdaa2e913b13148160b1152300c949bbb5f"},
+    {file = "numpy-1.21.3-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:3c09418a14471c7ae69ba682e2428cae5b4420a766659605566c0fa6987f6b7e"},
+    {file = "numpy-1.21.3-cp37-cp37m-win32.whl", hash = "sha256:90bec6a86b348b4559b6482e2b684db4a9a7eed1fa054b86115a48d58fbbf62a"},
+    {file = "numpy-1.21.3-cp37-cp37m-win_amd64.whl", hash = "sha256:043e83bfc274649c82a6f09836943e4a4aebe5e33656271c7dbf9621dd58b8ec"},
+    {file = "numpy-1.21.3-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:75621882d2230ab77fb6a03d4cbccd2038511491076e7964ef87306623aa5272"},
+    {file = "numpy-1.21.3-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:188031f833bbb623637e66006cf75e933e00e7231f67e2b45cf8189612bb5dc3"},
+    {file = "numpy-1.21.3-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:160ccc1bed3a8371bf0d760971f09bfe80a3e18646620e9ded0ad159d9749baa"},
+    {file = "numpy-1.21.3-cp38-cp38-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:29fb3dcd0468b7715f8ce2c0c2d9bbbaf5ae686334951343a41bd8d155c6ea27"},
+    {file = "numpy-1.21.3-cp38-cp38-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:32437f0b275c1d09d9c3add782516413e98cd7c09e6baf4715cbce781fc29912"},
+    {file = "numpy-1.21.3-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e606e6316911471c8d9b4618e082635cfe98876007556e89ce03d52ff5e8fcf0"},
+    {file = "numpy-1.21.3-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:a99a6b067e5190ac6d12005a4d85aa6227c5606fa93211f86b1dafb16233e57d"},
+    {file = "numpy-1.21.3-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:dde972a1e11bb7b702ed0e447953e7617723760f420decb97305e66fb4afc54f"},
+    {file = "numpy-1.21.3-cp38-cp38-win32.whl", hash = "sha256:fe52dbe47d9deb69b05084abd4b0df7abb39a3c51957c09f635520abd49b29dd"},
+    {file = "numpy-1.21.3-cp38-cp38-win_amd64.whl", hash = "sha256:75eb7cadc8da49302f5b659d40ba4f6d94d5045fbd9569c9d058e77b0514c9e4"},
+    {file = "numpy-1.21.3-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:2a6ee9620061b2a722749b391c0d80a0e2ae97290f1b32e28d5a362e21941ee4"},
+    {file = "numpy-1.21.3-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:5c4193f70f8069550a1788bd0cd3268ab7d3a2b70583dfe3b2e7f421e9aace06"},
+    {file = "numpy-1.21.3-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:28f15209fb535dd4c504a7762d3bc440779b0e37d50ed810ced209e5cea60d96"},
+    {file = "numpy-1.21.3-cp39-cp39-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:c6c2d535a7beb1f8790aaa98fd089ceab2e3dd7ca48aca0af7dc60e6ef93ffe1"},
+    {file = "numpy-1.21.3-cp39-cp39-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:bffa2eee3b87376cc6b31eee36d05349571c236d1de1175b804b348dc0941e3f"},
+    {file = "numpy-1.21.3-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc14e7519fab2a4ed87d31f99c31a3796e4e1fe63a86ebdd1c5a1ea78ebd5896"},
+    {file = "numpy-1.21.3-cp39-cp39-win32.whl", hash = "sha256:dd0482f3fc547f1b1b5d6a8b8e08f63fdc250c58ce688dedd8851e6e26cff0f3"},
+    {file = "numpy-1.21.3-cp39-cp39-win_amd64.whl", hash = "sha256:300321e3985c968e3ae7fbda187237b225f3ffe6528395a5b7a5407f73cf093e"},
+    {file = "numpy-1.21.3-pp37-pypy37_pp73-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:98339aa9911853f131de11010f6dd94c8cec254d3d1f7261528c3b3e3219f139"},
+    {file = "numpy-1.21.3.zip", hash = "sha256:63571bb7897a584ca3249c86dd01c10bcb5fe4296e3568b2e9c1a55356b6410e"},
+]
+onnxruntime = [
+    {file = "onnxruntime-1.9.0-cp36-cp36m-macosx_10_14_x86_64.whl", hash = "sha256:48f0fcf3c9aa6836584e64abe63fa7395c02066d3259bbdeb489b4d172e0127a"},
+    {file = "onnxruntime-1.9.0-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7f9d772a6330cb85e7723f84e357320a1603e3824a92aab4ef36fc3a41e64f16"},
+    {file = "onnxruntime-1.9.0-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3bdb861822a63404cca7b46dce86d48bbc21c906a4b4ed13969bc89763ac7f96"},
+    {file = "onnxruntime-1.9.0-cp36-cp36m-win32.whl", hash = "sha256:e3f8f7d5d4d66e3a4a2b731a000d3142a53a5403e8814e68bbd659514e815899"},
+    {file = "onnxruntime-1.9.0-cp36-cp36m-win_amd64.whl", hash = "sha256:cf3edbc54bfe99a119d73cd65398a2ec68ae3af2557ab7e645976314a8d11aa1"},
+    {file = "onnxruntime-1.9.0-cp37-cp37m-macosx_10_14_x86_64.whl", hash = "sha256:c8ff9c914b2b1c3b022dedc199e3f971e340d8923a1ef42d66530508fa367bf6"},
+    {file = "onnxruntime-1.9.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9d03ff4a2717c4149acc7c649fd66a67e81ec44c9e6e2a00df1d6e9ca843f1b7"},
+    {file = "onnxruntime-1.9.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a8a2315e2244ac371742f6e30da5367c680c3e84c31e291a35f8ddfab09c3c82"},
+    {file = "onnxruntime-1.9.0-cp37-cp37m-win32.whl", hash = "sha256:fa927b1825f2851c0c8f3948515a56d76cb0686da9acd1d6f8fafe552c8d8fec"},
+    {file = "onnxruntime-1.9.0-cp37-cp37m-win_amd64.whl", hash = "sha256:9ccaf6a0365f2b86efe21681416b8cfe97f084a7d53bd1cf2bf889a0aef2b0d3"},
+    {file = "onnxruntime-1.9.0-cp38-cp38-macosx_10_14_x86_64.whl", hash = "sha256:d20ce3448babe89a77cc9d357730767deb3617e36439bddcd006f28abc72b416"},
+    {file = "onnxruntime-1.9.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3fd1d6647245aa38e1099cfd355d84e807de5350d5216e84ceefd91c64ce243d"},
+    {file = "onnxruntime-1.9.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5f00620fc0f51bc4d90ae6d96ceb4b6538e3bd1e328178104118ac672f37c40d"},
+    {file = "onnxruntime-1.9.0-cp38-cp38-win32.whl", hash = "sha256:e1c1fe3f7d960eeffc02a5f196d85529254eefd59cbeecd8abee0a9467b5c2d8"},
+    {file = "onnxruntime-1.9.0-cp38-cp38-win_amd64.whl", hash = "sha256:9002214af1b2317ab3a63a2f045f7d1363c207e661d475a877aa6499ca09d606"},
+    {file = "onnxruntime-1.9.0-cp39-cp39-macosx_10_14_x86_64.whl", hash = "sha256:7339cef9b918b88f1fec8109cfa0a8416f119c5968d00300a9186847d86e35de"},
+    {file = "onnxruntime-1.9.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cdec8538eb59e63a376d0677f7ec043ceb597d52ee88f1f7e250928893a0de7f"},
+    {file = "onnxruntime-1.9.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ab62b29429e0e62c11478b2a8a3af2646531fba7800736e8b201d8baa50b43a8"},
+    {file = "onnxruntime-1.9.0-cp39-cp39-win32.whl", hash = "sha256:4aee9a893f93637341fd0e6b56fa3ab1c430d718d08d79a358603297a1575ad9"},
+    {file = "onnxruntime-1.9.0-cp39-cp39-win_amd64.whl", hash = "sha256:bdf1327932227383b04093a51266474b2703b3fcf9c0f6f11c652d9652b76a5c"},
+]
+opencv-python = [
+    {file = "opencv-python-4.5.4.58.tar.gz", hash = "sha256:48288428f407bacba5f73d460feb4a1ecafe87db3d7cfc0730a49fb32f589bbf"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:0eba0bfe62c48a02a5af3a0944e872c99f57f98653bed14d51c6991a58f9e1d1"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-macosx_11_0_x86_64.whl", hash = "sha256:9bcca50c5444b5cfb01624666b69f91ba8f2d2bf4ef37b111697aafdeb81c99f"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:8f7886acabaebf0361bd3dbccaa0d08e3f65ab13b7c739eb11e028f01ad13582"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-manylinux2014_x86_64.whl", hash = "sha256:d4b1d0b98ee72ba5dd720166790fc93ce459281e138ee79b0d41420b3da52b2e"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-win32.whl", hash = "sha256:69a78e40a374ac14e4bf15a13dbb6c30fd2fbd5fcd3674d020a31b88861d5aaf"},
+    {file = "opencv_python-4.5.4.58-cp310-cp310-win_amd64.whl", hash = "sha256:315c357522b6310ef7a0718d9f0c5d3110e59c19140705499a3c29bdd8c0124f"},
+    {file = "opencv_python-4.5.4.58-cp36-cp36m-macosx_10_15_x86_64.whl", hash = "sha256:887a61097092dc0bf23fa24646dbc8cfeeb753649cb28a3782a93a6879e3b7d2"},
+    {file = "opencv_python-4.5.4.58-cp36-cp36m-manylinux2014_aarch64.whl", hash = "sha256:22bcc3153a7d4f95aff79457eef81ef5e40ab1851b189e014412b5e9fbee2573"},
+    {file = "opencv_python-4.5.4.58-cp36-cp36m-manylinux2014_x86_64.whl", hash = "sha256:92e9b2261ec764229c948d77fe0d922ee033348ca6519939b87861016c1614b3"},
+    {file = "opencv_python-4.5.4.58-cp36-cp36m-win32.whl", hash = "sha256:0d6249a49122a78afc6685ddb1377a87e46414ae61c84535c4c6024397f1f3e8"},
+    {file = "opencv_python-4.5.4.58-cp36-cp36m-win_amd64.whl", hash = "sha256:eaa144013b597e4dcabc8d8230edfe810319de01b5609556d415a20e2b707547"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-macosx_10_15_x86_64.whl", hash = "sha256:26feeeb280de179f5dbb8976ebf7ceb836bd263973cb5daec8ca36e8ef7b5773"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-macosx_11_0_arm64.whl", hash = "sha256:4a13381bdfc0fb4b080efcc27c46561d0bd752f126226e9f19aa9cbcf6677f40"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-manylinux2014_aarch64.whl", hash = "sha256:ac852fcaac93439f2f7116ddffdc23fd366c872200ade2272446f9898180cecb"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-manylinux2014_x86_64.whl", hash = "sha256:02872e0a9358526646d691f390143e9c21109c210095314abaa0641211cda077"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-win32.whl", hash = "sha256:6b87bab220d17e03eeedbcc6652d9d7e7bb09886dbd0f810310697a948b4c6fd"},
+    {file = "opencv_python-4.5.4.58-cp37-cp37m-win_amd64.whl", hash = "sha256:a2a7f09b8843b85f3e1b02c5ea3ddc0cb9f5ad9698380109b37069ee8db7746d"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-macosx_10_15_x86_64.whl", hash = "sha256:c44f5c51e92322ed832607204249c190764dec6cf29e8ba6d679b10326be1c1b"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:9b2c198af083a693d42a82bddc4d1f7e6bb02c64192ff7fac1fd1d43a8cf1be6"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-manylinux2014_aarch64.whl", hash = "sha256:637f4d3ad81bd27f273ede4c5fa6c26afb85c097c9715baf107cc270e37f5fea"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-manylinux2014_x86_64.whl", hash = "sha256:2fff48a641a74d1def31c1e88f9e5ce50ba4d0f87d085dfbf8bc844e12f6cd54"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-win32.whl", hash = "sha256:8ddf4dcd8199209e33f21deb0c6d8ab62b21802816bba895fefc346b6d2e522d"},
+    {file = "opencv_python-4.5.4.58-cp38-cp38-win_amd64.whl", hash = "sha256:085c5fcf5a6479c34aca3fd0f59055e704083d6a44009d6583c675ff1a5a0625"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-macosx_10_15_x86_64.whl", hash = "sha256:4abe9c4fb6fe16daa9fcdd68b5357d3530431341aa655203f8e84f394e1fe6d4"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:4b614fbd81aeda53ce28e645aaee18fda7c7f2a48eb7f1a70a7c6c3427946342"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:215bdf069847d4e3b0447a34e9eb4046dd4ca523d41fe4381c1c55f6704fd0dc"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc34cdbfbab463750713118c8259a5d364547adab8ed91e94ba888349f33590a"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-win32.whl", hash = "sha256:9998ce60884f3cda074f02b56d2b57ee6bd863e2ddba132da2b0af3b9487d584"},
+    {file = "opencv_python-4.5.4.58-cp39-cp39-win_amd64.whl", hash = "sha256:5370a11757fbe94b176771269aff599f4da8676c2a672b13bcbca043f2e3eea8"},
+]
+pillow = [
+    {file = "Pillow-8.4.0-cp310-cp310-macosx_10_10_universal2.whl", hash = "sha256:81f8d5c81e483a9442d72d182e1fb6dcb9723f289a57e8030811bac9ea3fef8d"},
+    {file = "Pillow-8.4.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:3f97cfb1e5a392d75dd8b9fd274d205404729923840ca94ca45a0af57e13dbe6"},
+    {file = "Pillow-8.4.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:eb9fc393f3c61f9054e1ed26e6fe912c7321af2f41ff49d3f83d05bacf22cc78"},
+    {file = "Pillow-8.4.0-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d82cdb63100ef5eedb8391732375e6d05993b765f72cb34311fab92103314649"},
+    {file = "Pillow-8.4.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:62cc1afda735a8d109007164714e73771b499768b9bb5afcbbee9d0ff374b43f"},
+    {file = "Pillow-8.4.0-cp310-cp310-win32.whl", hash = "sha256:e3dacecfbeec9a33e932f00c6cd7996e62f53ad46fbe677577394aaa90ee419a"},
+    {file = "Pillow-8.4.0-cp310-cp310-win_amd64.whl", hash = "sha256:620582db2a85b2df5f8a82ddeb52116560d7e5e6b055095f04ad828d1b0baa39"},
+    {file = "Pillow-8.4.0-cp36-cp36m-macosx_10_10_x86_64.whl", hash = "sha256:1bc723b434fbc4ab50bb68e11e93ce5fb69866ad621e3c2c9bdb0cd70e345f55"},
+    {file = "Pillow-8.4.0-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:72cbcfd54df6caf85cc35264c77ede902452d6df41166010262374155947460c"},
+    {file = "Pillow-8.4.0-cp36-cp36m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:70ad9e5c6cb9b8487280a02c0ad8a51581dcbbe8484ce058477692a27c151c0a"},
+    {file = "Pillow-8.4.0-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:25a49dc2e2f74e65efaa32b153527fc5ac98508d502fa46e74fa4fd678ed6645"},
+    {file = "Pillow-8.4.0-cp36-cp36m-win32.whl", hash = "sha256:93ce9e955cc95959df98505e4608ad98281fff037350d8c2671c9aa86bcf10a9"},
+    {file = "Pillow-8.4.0-cp36-cp36m-win_amd64.whl", hash = "sha256:2e4440b8f00f504ee4b53fe30f4e381aae30b0568193be305256b1462216feff"},
+    {file = "Pillow-8.4.0-cp37-cp37m-macosx_10_10_x86_64.whl", hash = "sha256:8c803ac3c28bbc53763e6825746f05cc407b20e4a69d0122e526a582e3b5e153"},
+    {file = "Pillow-8.4.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c8a17b5d948f4ceeceb66384727dde11b240736fddeda54ca740b9b8b1556b29"},
+    {file = "Pillow-8.4.0-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:1394a6ad5abc838c5cd8a92c5a07535648cdf6d09e8e2d6df916dfa9ea86ead8"},
+    {file = "Pillow-8.4.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:792e5c12376594bfcb986ebf3855aa4b7c225754e9a9521298e460e92fb4a488"},
+    {file = "Pillow-8.4.0-cp37-cp37m-win32.whl", hash = "sha256:d99ec152570e4196772e7a8e4ba5320d2d27bf22fdf11743dd882936ed64305b"},
+    {file = "Pillow-8.4.0-cp37-cp37m-win_amd64.whl", hash = "sha256:7b7017b61bbcdd7f6363aeceb881e23c46583739cb69a3ab39cb384f6ec82e5b"},
+    {file = "Pillow-8.4.0-cp38-cp38-macosx_10_10_x86_64.whl", hash = "sha256:d89363f02658e253dbd171f7c3716a5d340a24ee82d38aab9183f7fdf0cdca49"},
+    {file = "Pillow-8.4.0-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:0a0956fdc5defc34462bb1c765ee88d933239f9a94bc37d132004775241a7585"},
+    {file = "Pillow-8.4.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5b7bb9de00197fb4261825c15551adf7605cf14a80badf1761d61e59da347779"},
+    {file = "Pillow-8.4.0-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:72b9e656e340447f827885b8d7a15fc8c4e68d410dc2297ef6787eec0f0ea409"},
+    {file = "Pillow-8.4.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a5a4532a12314149d8b4e4ad8ff09dde7427731fcfa5917ff16d0291f13609df"},
+    {file = "Pillow-8.4.0-cp38-cp38-win32.whl", hash = "sha256:82aafa8d5eb68c8463b6e9baeb4f19043bb31fefc03eb7b216b51e6a9981ae09"},
+    {file = "Pillow-8.4.0-cp38-cp38-win_amd64.whl", hash = "sha256:066f3999cb3b070a95c3652712cffa1a748cd02d60ad7b4e485c3748a04d9d76"},
+    {file = "Pillow-8.4.0-cp39-cp39-macosx_10_10_x86_64.whl", hash = "sha256:5503c86916d27c2e101b7f71c2ae2cddba01a2cf55b8395b0255fd33fa4d1f1a"},
+    {file = "Pillow-8.4.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:4acc0985ddf39d1bc969a9220b51d94ed51695d455c228d8ac29fcdb25810e6e"},
+    {file = "Pillow-8.4.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0b052a619a8bfcf26bd8b3f48f45283f9e977890263e4571f2393ed8898d331b"},
+    {file = "Pillow-8.4.0-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:493cb4e415f44cd601fcec11c99836f707bb714ab03f5ed46ac25713baf0ff20"},
+    {file = "Pillow-8.4.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b8831cb7332eda5dc89b21a7bce7ef6ad305548820595033a4b03cf3091235ed"},
+    {file = "Pillow-8.4.0-cp39-cp39-win32.whl", hash = "sha256:5e9ac5f66616b87d4da618a20ab0a38324dbe88d8a39b55be8964eb520021e02"},
+    {file = "Pillow-8.4.0-cp39-cp39-win_amd64.whl", hash = "sha256:3eb1ce5f65908556c2d8685a8f0a6e989d887ec4057326f6c22b24e8a172c66b"},
+    {file = "Pillow-8.4.0-pp36-pypy36_pp73-macosx_10_10_x86_64.whl", hash = "sha256:ddc4d832a0f0b4c52fff973a0d44b6c99839a9d016fe4e6a1cb8f3eea96479c2"},
+    {file = "Pillow-8.4.0-pp36-pypy36_pp73-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9a3e5ddc44c14042f0844b8cf7d2cd455f6cc80fd7f5eefbe657292cf601d9ad"},
+    {file = "Pillow-8.4.0-pp36-pypy36_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c70e94281588ef053ae8998039610dbd71bc509e4acbc77ab59d7d2937b10698"},
+    {file = "Pillow-8.4.0-pp37-pypy37_pp73-macosx_10_10_x86_64.whl", hash = "sha256:3862b7256046fcd950618ed22d1d60b842e3a40a48236a5498746f21189afbbc"},
+    {file = "Pillow-8.4.0-pp37-pypy37_pp73-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a4901622493f88b1a29bd30ec1a2f683782e57c3c16a2dbc7f2595ba01f639df"},
+    {file = "Pillow-8.4.0-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:84c471a734240653a0ec91dec0996696eea227eafe72a33bd06c92697728046b"},
+    {file = "Pillow-8.4.0-pp37-pypy37_pp73-win_amd64.whl", hash = "sha256:244cf3b97802c34c41905d22810846802a3329ddcb93ccc432870243211c79fc"},
+    {file = "Pillow-8.4.0.tar.gz", hash = "sha256:b8e2f83c56e141920c39464b852de3719dfbfb6e3c99a2d8da0edf4fb33176ed"},
+]
+protobuf = [
+    {file = "protobuf-3.19.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d80f80eb175bf5f1169139c2e0c5ada98b1c098e2b3c3736667f28cbbea39fc8"},
+    {file = "protobuf-3.19.1-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:a529e7df52204565bcd33738a7a5f288f3d2d37d86caa5d78c458fa5fabbd54d"},
+    {file = "protobuf-3.19.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:28ccea56d4dc38d35cd70c43c2da2f40ac0be0a355ef882242e8586c6d66666f"},
+    {file = "protobuf-3.19.1-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:8b30a7de128c46b5ecb343917d9fa737612a6e8280f440874e5cc2ba0d79b8f6"},
+    {file = "protobuf-3.19.1-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5935c8ce02e3d89c7900140a8a42b35bc037ec07a6aeb61cc108be8d3c9438a6"},
+    {file = "protobuf-3.19.1-cp36-cp36m-win32.whl", hash = "sha256:74f33edeb4f3b7ed13d567881da8e5a92a72b36495d57d696c2ea1ae0cfee80c"},
+    {file = "protobuf-3.19.1-cp36-cp36m-win_amd64.whl", hash = "sha256:038daf4fa38a7e818dd61f51f22588d61755160a98db087a046f80d66b855942"},
+    {file = "protobuf-3.19.1-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:8e51561d72efd5bd5c91490af1f13e32bcba8dab4643761eb7de3ce18e64a853"},
+    {file = "protobuf-3.19.1-cp37-cp37m-manylinux2014_aarch64.whl", hash = "sha256:6e8ea9173403219239cdfd8d946ed101f2ab6ecc025b0fda0c6c713c35c9981d"},
+    {file = "protobuf-3.19.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:db3532d9f7a6ebbe2392041350437953b6d7a792de10e629c1e4f5a6b1fe1ac6"},
+    {file = "protobuf-3.19.1-cp37-cp37m-win32.whl", hash = "sha256:615b426a177780ce381ecd212edc1e0f70db8557ed72560b82096bd36b01bc04"},
+    {file = "protobuf-3.19.1-cp37-cp37m-win_amd64.whl", hash = "sha256:d8919368410110633717c406ab5c97e8df5ce93020cfcf3012834f28b1fab1ea"},
+    {file = "protobuf-3.19.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:71b0250b0cfb738442d60cab68abc166de43411f2a4f791d31378590bfb71bd7"},
+    {file = "protobuf-3.19.1-cp38-cp38-manylinux2014_aarch64.whl", hash = "sha256:3cd0458870ea7d1c58e948ac8078f6ba8a7ecc44a57e03032ed066c5bb318089"},
+    {file = "protobuf-3.19.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:655264ed0d0efe47a523e2255fc1106a22f6faab7cc46cfe99b5bae085c2a13e"},
+    {file = "protobuf-3.19.1-cp38-cp38-win32.whl", hash = "sha256:b691d996c6d0984947c4cf8b7ae2fe372d99b32821d0584f0b90277aa36982d3"},
+    {file = "protobuf-3.19.1-cp38-cp38-win_amd64.whl", hash = "sha256:e7e8d2c20921f8da0dea277dfefc6abac05903ceac8e72839b2da519db69206b"},
+    {file = "protobuf-3.19.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:fd390367fc211cc0ffcf3a9e149dfeca78fecc62adb911371db0cec5c8b7472d"},
+    {file = "protobuf-3.19.1-cp39-cp39-manylinux2014_aarch64.whl", hash = "sha256:d83e1ef8cb74009bebee3e61cc84b1c9cd04935b72bca0cbc83217d140424995"},
+    {file = "protobuf-3.19.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:36d90676d6f426718463fe382ec6274909337ca6319d375eebd2044e6c6ac560"},
+    {file = "protobuf-3.19.1-cp39-cp39-win32.whl", hash = "sha256:e7b24c11df36ee8e0c085e5b0dc560289e4b58804746fb487287dda51410f1e2"},
+    {file = "protobuf-3.19.1-cp39-cp39-win_amd64.whl", hash = "sha256:77d2fadcf369b3f22859ab25bd12bb8e98fb11e05d9ff9b7cd45b711c719c002"},
+    {file = "protobuf-3.19.1-py2.py3-none-any.whl", hash = "sha256:e813b1c9006b6399308e917ac5d298f345d95bb31f46f02b60cd92970a9afa17"},
+    {file = "protobuf-3.19.1.tar.gz", hash = "sha256:62a8e4baa9cb9e064eb62d1002eca820857ab2138440cb4b3ea4243830f94ca7"},
+]

pyproject.toml

@@ -0,0 +1,19 @@
+[tool.poetry]
+name = "openseeface"
+version = "1.20.4"
+description = ""
+authors = ["emilianavt"]
+license = "BSD-2-Clause"
+
+[tool.poetry.dependencies]
+python = ">=3.7,<3.11"
+numpy = "^1.21.3"
+opencv-python = "^4.5.4"
+Pillow = "^8.4.0"
+onnxruntime = "^1.9.0"
+
+[tool.poetry.dev-dependencies]
+
+[build-system]
+requires = ["poetry-core>=1.0.0"]
+build-backend = "poetry.core.masonry.api"

retinaface.py

@@ -66,7 +66,8 @@ class RetinaFaceDetector():
         options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
         options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         options.log_severity_level = 3
-        self.session = onnxruntime.InferenceSession(model_path, sess_options=options)
+        providersList = onnxruntime.capi._pybind_state.get_available_providers()
+        self.session = onnxruntime.InferenceSession(model_path, sess_options=options, providers=providersList)
         self.res_w, self.res_h = res
         with open(json_path, "r") as prior_file:
             self.priorbox = np.array(json.loads(prior_file.read()))

run.bat

@@ -2,12 +2,15 @@
 
 facetracker -l 1
 
+echo Make sure that nothing is accessing your camera before you proceed.
 
 set /p cameraNum=Select your camera from the list above and enter the corresponding number:
-set /p width=Select the width:
+
-set /p height=Select the height:
+facetracker -a %cameraNum%
+
+set /p dcaps=Select your camera mode or -1 for default settings:
 set /p fps=Select the FPS:
 
-facetracker -c %cameraNum% -W %width% -H %height% -F %fps% -v 3 -P 1 --discard-after 0 --scan-every 0 --no-3d-adapt 1 --max-feature-updates 900
+facetracker -c %cameraNum% -F %fps% -D %dcaps% -v 3 -P 1 --discard-after 0 --scan-every 0 --no-3d-adapt 1 --max-feature-updates 900
 
 pause

tracker.py

@@ -7,6 +7,7 @@ import time
 import queue
 import threading
 import copy
+from similaritytransform import SimilarityTransform
 from retinaface import RetinaFaceDetector
 from remedian import remedian
 
@@ -49,7 +50,7 @@ def compensate(p1, p2):
 def rotate_image(image, a, center):
     (h, w) = image.shape[:2]
     a = np.rad2deg(a)
-    M = cv2.getRotationMatrix2D((center[0], center[1]), a, 1.0)
+    M = cv2.getRotationMatrix2D((float(center[0]), float(center[1])), a, 1.0)
     rotated = cv2.warpAffine(image, M, (w, h))
     return rotated
 
@@ -110,6 +111,10 @@ def logit(p, factor=16.0):
     p = p/(1-p)
     return float(np.log(p)) / float(factor)
 
+def logit_arr(p, factor=16.0):
+    p = np.clip(p, 0.0000001, 0.9999999)
+    return np.log(p / (1 - p)) / float(factor)
+
 def matrix_to_quaternion(m):
     t = 0.0
     q = [0.0, 0.0, 0, 0.0]
@@ -134,7 +139,10 @@ def worker_thread(session, frame, input, crop_info, queue, input_name, idx, trac
     output = session.run([], {input_name: input})[0]
     conf, lms = tracker.landmarks(output[0], crop_info)
     if conf > tracker.threshold:
-        eye_state = tracker.get_eye_state(frame, lms, single=True)
+        try:
+            eye_state = tracker.get_eye_state(frame, lms)
+        except:
+            eye_state = [(1.0, 0.0, 0.0, 0.0), (1.0, 0.0, 0.0, 0.0)]
         queue.put((session, conf, (lms, eye_state), crop_info, idx))
     else:
         queue.put((session,))
@@ -247,7 +255,7 @@ class FeatureExtractor():
             aligned_pts.append(np.array(rotate(a, pt, alpha)))
         return alpha, np.array(aligned_pts)
 
-    def update(self, pts):
+    def update(self, pts, full=True):
         features = {}
         now = time.perf_counter()
 
@@ -255,45 +263,59 @@ class FeatureExtractor():
         norm_distance_y = np.mean([pts[27, 1] - pts[28, 1], pts[28, 1] - pts[29, 1], pts[29, 1] - pts[30, 1]])
 
         a1, f_pts = self.align_points(pts[42], pts[45], pts[[43, 44, 47, 46]])
-        f = np.clip((np.mean([f_pts[0,1], f_pts[1,1]]) - np.mean([f_pts[2,1], f_pts[3,1]])) / norm_distance_y, 0, None)
+        f = abs((np.mean([f_pts[0,1], f_pts[1,1]]) - np.mean([f_pts[2,1], f_pts[3,1]])) / norm_distance_y)
         features["eye_l"] = self.eye_l.update(f, now)
 
         a2, f_pts = self.align_points(pts[36], pts[39], pts[[37, 38, 41, 40]])
-        f = np.clip((np.mean([f_pts[0,1], f_pts[1,1]]) - np.mean([f_pts[2,1], f_pts[3,1]])) / norm_distance_y, 0, None)
+        f = abs((np.mean([f_pts[0,1], f_pts[1,1]]) - np.mean([f_pts[2,1], f_pts[3,1]])) / norm_distance_y)
         features["eye_r"] = self.eye_r.update(f, now)
 
-        a3, _ = self.align_points(pts[0], pts[16], [])
+        if full:
-        a4, _ = self.align_points(pts[31], pts[35], [])
+            a3, _ = self.align_points(pts[0], pts[16], [])
-        norm_angle = np.mean(list(map(np.rad2deg, [a1, a2, a3, a4])))
+            a4, _ = self.align_points(pts[31], pts[35], [])
+            norm_angle = np.mean(list(map(np.rad2deg, [a1, a2, a3, a4])))
+
+            a, f_pts = self.align_points(pts[22], pts[26], pts[[22, 23, 24, 25, 26]])
+            features["eyebrow_steepness_l"] = self.eyebrow_steepness_l.update(-np.rad2deg(a) - norm_angle, now)
+            f = np.max(np.abs(np.array(f_pts[1:4]) - f_pts[0, 1])) / norm_distance_y
+            features["eyebrow_quirk_l"] = self.eyebrow_quirk_l.update(f, now)
+
+            a, f_pts = self.align_points(pts[17], pts[21], pts[[17, 18, 19, 20, 21]])
+            features["eyebrow_steepness_r"] = self.eyebrow_steepness_r.update(np.rad2deg(a) - norm_angle, now)
+            f = np.max(np.abs(np.array(f_pts[1:4]) - f_pts[0, 1])) / norm_distance_y
+            features["eyebrow_quirk_r"] = self.eyebrow_quirk_r.update(f, now)
+        else:
+            features["eyebrow_steepness_l"] = 0.
+            features["eyebrow_steepness_r"] = 0.
+            features["eyebrow_quirk_l"] = 0.
+            features["eyebrow_quirk_r"] = 0.
 
-        a, f_pts = self.align_points(pts[22], pts[26], pts[[22, 23, 24, 25, 26]])
-        features["eyebrow_steepness_l"] = self.eyebrow_steepness_l.update(-np.rad2deg(a) - norm_angle, now)
         f = (np.mean([pts[22, 1], pts[26, 1]]) - pts[27, 1]) / norm_distance_y
         features["eyebrow_updown_l"] = self.eyebrow_updown_l.update(f, now)
-        f = np.max(np.abs(np.array(f_pts[1:4]) - f_pts[0, 1])) / norm_distance_y
-        features["eyebrow_quirk_l"] = self.eyebrow_quirk_l.update(f, now)
 
-        a, f_pts = self.align_points(pts[17], pts[21], pts[[17, 18, 19, 20, 21]])
-        features["eyebrow_steepness_r"] = self.eyebrow_steepness_r.update(np.rad2deg(a) - norm_angle, now)
         f = (np.mean([pts[17, 1], pts[21, 1]]) - pts[27, 1]) / norm_distance_y
         features["eyebrow_updown_r"] = self.eyebrow_updown_r.update(f, now)
-        f = np.max(np.abs(np.array(f_pts[1:4]) - f_pts[0, 1])) / norm_distance_y
-        features["eyebrow_quirk_r"] = self.eyebrow_quirk_r.update(f, now)
 
         upper_mouth_line = np.mean([pts[49, 1], pts[50, 1], pts[51, 1]])
         center_line = np.mean([pts[50, 0], pts[60, 0], pts[27, 0], pts[30, 0], pts[64, 0], pts[55, 0]])
 
         f = (upper_mouth_line - pts[62, 1]) / norm_distance_y
         features["mouth_corner_updown_l"] = self.mouth_corner_updown_l.update(f, now)
-        f = abs(center_line - pts[62, 0]) / norm_distance_x
+        if full:
-        features["mouth_corner_inout_l"] = self.mouth_corner_inout_l.update(f, now)
+            f = abs(center_line - pts[62, 0]) / norm_distance_x
+            features["mouth_corner_inout_l"] = self.mouth_corner_inout_l.update(f, now)
+        else:
+            features["mouth_corner_inout_l"] = 0.
 
         f = (upper_mouth_line - pts[58, 1]) / norm_distance_y
         features["mouth_corner_updown_r"] = self.mouth_corner_updown_r.update(f, now)
-        f = abs(center_line - pts[58, 0]) / norm_distance_x
+        if full:
-        features["mouth_corner_inout_r"] = self.mouth_corner_inout_r.update(f, now)
+            f = abs(center_line - pts[58, 0]) / norm_distance_x
+            features["mouth_corner_inout_r"] = self.mouth_corner_inout_r.update(f, now)
+        else:
+            features["mouth_corner_inout_r"] = 0.
 
-        f = (np.mean([pts[59, 1], pts[60, 1], pts[61, 1]]) - np.mean([pts[65, 1], pts[64, 1], pts[63, 1]])) / norm_distance_y
+        f = abs(np.mean(pts[[59,60,61], 1], axis=0) - np.mean(pts[[63,64,65], 1], axis=0)) / norm_distance_y
         features["mouth_open"] = self.mouth_open.update(f, now)
 
         f = abs(pts[58, 0] - pts[62, 0]) / norm_distance_x
@@ -306,8 +328,10 @@ class FaceInfo():
         self.id = id
         self.frame_count = -1
         self.tracker = tracker
-        self.contour_pts = [0,1,8,15,16,27,28,29,30,31,32,33,34,35,36,39,42,45]
+        self.contour_pts = [0,1,8,15,16,27,28,29,30,31,32,33,34,35]
         self.face_3d = copy.copy(self.tracker.face_3d)
+        if self.tracker.model_type == -1:
+            self.contour_pts = [0,2,8,14,16,27,30,33]
         self.reset()
         self.alive = False
         self.coord = None
@@ -377,7 +401,7 @@ class FaceInfo():
         if self.conf < 0.4 or self.pnp_error > 300:
             return
 
-        if not self.tracker.static_model:
+        if self.tracker.model_type != -1 and not self.tracker.static_model:
             max_runs = 1
             eligible = np.delete(np.arange(0, 66), [30])
             changed_any = False
@@ -449,70 +473,93 @@ class FaceInfo():
                 self.update_contour()
 
         self.pts_3d = self.normalize_pts3d(self.pts_3d)
-        self.current_features = self.features.update(self.pts_3d[:, 0:2])
+        if self.tracker.feature_level == 2:
-        self.eye_blink = []
+            self.current_features = self.features.update(self.pts_3d[:, 0:2])
-        self.eye_blink.append(1 - min(max(0, -self.current_features["eye_r"]), 1))
+            self.eye_blink = []
-        self.eye_blink.append(1 - min(max(0, -self.current_features["eye_l"]), 1))
+            self.eye_blink.append(1 - min(max(0, -self.current_features["eye_r"]), 1))
+            self.eye_blink.append(1 - min(max(0, -self.current_features["eye_l"]), 1))
+        elif self.tracker.feature_level == 1:
+            self.current_features = self.features.update(self.pts_3d[:, 0:2], False)
+            self.eye_blink = []
+            self.eye_blink.append(1 - min(max(0, -self.current_features["eye_r"]), 1))
+            self.eye_blink.append(1 - min(max(0, -self.current_features["eye_l"]), 1))
 
+def get_model_base_path(model_dir):
+    model_base_path = resolve(os.path.join("models"))
+    if model_dir is None:
+        if not os.path.exists(model_base_path):
+            model_base_path = resolve(os.path.join("..", "models"))
+    else:
+        model_base_path = model_dir
+    return model_base_path
 
 class Tracker():
-    def __init__(self, width, height, model_type=3, threshold=0.6, max_faces=1, discard_after=5, scan_every=3, bbox_growth=0.0, max_threads=4, silent=False, model_dir=None, no_gaze=False, use_retinaface=False, max_feature_updates=0, static_model=False):
+    def __init__(self, width, height, model_type=3, detection_threshold=0.6, threshold=None, max_faces=1, discard_after=5, scan_every=3, bbox_growth=0.0, max_threads=4, silent=False, model_dir=None, no_gaze=False, use_retinaface=False, max_feature_updates=0, static_model=False, feature_level=2, try_hard=False):
         options = onnxruntime.SessionOptions()
         options.inter_op_num_threads = 1
-        options.intra_op_num_threads = max(max_threads,4)
+        options.intra_op_num_threads = min(max_threads,4)
         options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
         options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         options.log_severity_level = 3
         self.model_type = model_type
         self.models = [
-            "mnv3_opt_very_fast.onnx",
+            "lm_model0_opt.onnx",
-            "mnv3_opt_fast.onnx",
+            "lm_model1_opt.onnx",
-            "mnv3_opt_medium.onnx",
+            "lm_model2_opt.onnx",
-            "mnv3_opt_b.onnx"
+            "lm_model3_opt.onnx",
+            "lm_model4_opt.onnx"
         ]
-        model = self.models[self.model_type]
+        model = "lm_modelT_opt.onnx"
-        model_base_path = resolve(os.path.join("models"))
+        if model_type >= 0:
-        if model_dir is None:
+            model = self.models[self.model_type]
-            if not os.path.exists(model_base_path):
+        if model_type == -2:
-                model_base_path = resolve(os.path.join("..", "models"))
+            model = "lm_modelV_opt.onnx"
-        else:
+        if model_type == -3:
-            model_base_path = model_dir
+            model = "lm_modelU_opt.onnx"
+        model_base_path = get_model_base_path(model_dir)
+
+        if threshold is None:
+            threshold = 0.6
+            if model_type < 0:
+                threshold = 0.87
 
         self.retinaface = RetinaFaceDetector(model_path=os.path.join(model_base_path, "retinaface_640x640_opt.onnx"), json_path=os.path.join(model_base_path, "priorbox_640x640.json"), threads=max(max_threads,4), top_k=max_faces, res=(640, 640))
         self.retinaface_scan = RetinaFaceDetector(model_path=os.path.join(model_base_path, "retinaface_640x640_opt.onnx"), json_path=os.path.join(model_base_path, "priorbox_640x640.json"), threads=2, top_k=max_faces, res=(640, 640))
         self.use_retinaface = use_retinaface
 
+        # OTR 1.9 and later requires specifying providers
+        # explicitly as an argument in InferenceSession()
+        providersList = onnxruntime.capi._pybind_state.get_available_providers()
+
         # Single face instance with multiple threads
-        self.session = onnxruntime.InferenceSession(os.path.join(model_base_path, model), sess_options=options)
+        self.session = onnxruntime.InferenceSession(os.path.join(model_base_path, model), sess_options=options, providers=providersList)
 
         # Multiple faces with single threads
         self.sessions = []
-        self.max_workers = min(max_threads, max_faces)
+        self.max_workers = max(min(max_threads, max_faces), 1)
         extra_threads = max_threads % self.max_workers
         for i in range(self.max_workers):
             options = onnxruntime.SessionOptions()
             options.inter_op_num_threads = 1
-            options.intra_op_num_threads = max_threads // self.max_workers
+            options.intra_op_num_threads = min(max(max_threads // self.max_workers, 4), 1)
             if options.intra_op_num_threads < 1:
                 options.intra_op_num_threads = 1
             elif i < extra_threads:
                 options.intra_op_num_threads += 1
             options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
             options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
-            self.sessions.append(onnxruntime.InferenceSession(os.path.join(model_base_path, model), sess_options=options))
+            self.sessions.append(onnxruntime.InferenceSession(os.path.join(model_base_path, model), sess_options=options, providers=providersList))
         self.input_name = self.session.get_inputs()[0].name
 
         options = onnxruntime.SessionOptions()
         options.inter_op_num_threads = 1
-        options.intra_op_num_threads = max(max_threads,4)
+        options.intra_op_num_threads = 1
         options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
         options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         options.log_severity_level = 3
-        self.gaze_model = onnxruntime.InferenceSession(os.path.join(model_base_path, "mnv3_gaze32_split_opt.onnx"), sess_options=options)
+        self.gaze_model = onnxruntime.InferenceSession(os.path.join(model_base_path, "mnv3_gaze32_split_opt.onnx"), sess_options=options, providers=providersList)
-        options.intra_op_num_threads = 1
-        self.gaze_model_single = onnxruntime.InferenceSession(os.path.join(model_base_path, "mnv3_gaze32_split_opt.onnx"), sess_options=options)
 
-        self.detection = onnxruntime.InferenceSession(os.path.join(model_base_path, "mnv3_detection_opt.onnx"), sess_options=options)
+        self.detection = onnxruntime.InferenceSession(os.path.join(model_base_path, "mnv3_detection_opt.onnx"), sess_options=options, providers=providersList)
         self.faces = []
 
         # Image normalization constants
@@ -521,6 +568,13 @@ class Tracker():
         self.mean = self.mean / self.std
         self.std = self.std * 255.0
 
+        self.mean = - self.mean
+        self.std = 1.0 / self.std
+        self.mean_32 = np.tile(self.mean, [32, 32, 1])
+        self.std_32 = np.tile(self.std, [32, 32, 1])
+        self.mean_224 = np.tile(self.mean, [224, 224, 1])
+        self.std_224 = np.tile(self.std, [224, 224, 1])
+
         # PnP solving
         self.face_3d = np.array([
             [ 0.4551769692672  ,  0.300895790030204, -0.764429433974752],
@@ -604,6 +658,7 @@ class Tracker():
         self.width = width
         self.height = height
         self.threshold = threshold
+        self.detection_threshold = detection_threshold
         self.max_faces = max_faces
         self.max_threads = max_threads
         self.discard = 0
@@ -613,17 +668,44 @@ class Tracker():
         self.scan_every = scan_every
         self.bbox_growth = bbox_growth
         self.silent = silent
+        self.try_hard = try_hard
+
         self.res = 224.
+        self.mean_res = self.mean_224
+        self.std_res = self.std_224
+        if model_type < 0:
+            self.res = 56.
+            self.mean_res = np.tile(self.mean, [56, 56, 1])
+            self.std_res = np.tile(self.std, [56, 56, 1])
+        if model_type < -1:
+            self.res = 112.
+            self.mean_res = np.tile(self.mean, [112, 112, 1])
+            self.std_res = np.tile(self.std, [112, 112, 1])
         self.res_i = int(self.res)
+        self.out_res = 27.
+        if model_type < 0:
+            self.out_res = 6.
+        if model_type < -1:
+            self.out_res = 13.
+        self.out_res_i = int(self.out_res) + 1
+        self.logit_factor = 16.
+        if model_type < 0:
+            self.logit_factor = 8.
+        if model_type < -1:
+            self.logit_factor = 16.
+
         self.no_gaze = no_gaze
         self.debug_gaze = False
+        self.feature_level = feature_level
+        if model_type == -1:
+            self.feature_level = min(feature_level, 1)
         self.max_feature_updates = max_feature_updates
         self.static_model = static_model
         self.face_info = [FaceInfo(id, self) for id in range(max_faces)]
         self.fail_count = 0
 
     def detect_faces(self, frame):
-        im = cv2.resize(frame, (224, 224), interpolation=cv2.INTER_LINEAR)[:,:,::-1] / self.std - self.mean
+        im = cv2.resize(frame, (224, 224), interpolation=cv2.INTER_LINEAR)[:,:,::-1] * self.std_224 + self.mean_224
         im = np.expand_dims(im, 0)
         im = np.transpose(im, (0,3,1,2))
         outputs, maxpool = self.detection.run([], {'input': im})
@@ -639,7 +721,7 @@ class Tracker():
             x *= 4
             y *= 4
             r *= 1.0
-            if c < self.threshold:
+            if c < self.detection_threshold:
                 break
             results.append((x - r, y - r, 2 * r, 2 * r * 1.0))
         results = np.array(results).astype(np.float32)
@@ -651,31 +733,41 @@ class Tracker():
     def landmarks(self, tensor, crop_info):
         crop_x1, crop_y1, scale_x, scale_y, _ = crop_info
         avg_conf = 0
-        lms = []
         res = self.res - 1
-        for i in range(0, 66):
+        c0, c1, c2 = 66, 132, 198
-            m = int(tensor[i].argmax())
+        if self.model_type == -1:
-            x = m // 28
+            c0, c1, c2 = 30, 60, 90
-            y = m % 28
+        t_main = tensor[0:c0].reshape((c0,self.out_res_i * self.out_res_i))
-            conf = float(tensor[i][x,y])
+        t_m = t_main.argmax(1)
-            avg_conf = avg_conf + conf
+        indices = np.expand_dims(t_m, 1)
-            off_x = res * ((1. * logit(tensor[66 + i][x, y])) - 0.0)
+        t_conf = np.take_along_axis(t_main, indices, 1).reshape((c0,))
-            off_y = res * ((1. * logit(tensor[66 * 2 + i][x, y])) - 0.0)
+        t_off_x = np.take_along_axis(tensor[c0:c1].reshape((c0,self.out_res_i * self.out_res_i)), indices, 1).reshape((c0,))
-            off_x = math.floor(off_x + 0.5)
+        t_off_y = np.take_along_axis(tensor[c1:c2].reshape((c0,self.out_res_i * self.out_res_i)), indices, 1).reshape((c0,))
-            off_y = math.floor(off_y + 0.5)
+        t_off_x = res * logit_arr(t_off_x, self.logit_factor)
-            lm_x = crop_y1 + scale_y * (res * (float(x) / 27.) + off_x)
+        t_off_y = res * logit_arr(t_off_y, self.logit_factor)
-            lm_y = crop_x1 + scale_x * (res * (float(y) / 27.) + off_y)
+        t_x = crop_y1 + scale_y * (res * np.floor(t_m / self.out_res_i) / self.out_res + t_off_x)
-            lms.append((lm_x,lm_y,conf))
+        t_y = crop_x1 + scale_x * (res * np.floor(np.mod(t_m, self.out_res_i)) / self.out_res + t_off_y)
-        avg_conf = avg_conf / 66.
+        avg_conf = np.average(t_conf)
+        lms = np.stack([t_x, t_y, t_conf], 1)
+        lms[np.isnan(lms).any(axis=1)] = np.array([0.,0.,0.], dtype=np.float32)
+        if self.model_type == -1:
+            lms = lms[[0,0,1,1,1,2,2,2,3,3,3,4,4,4,5,5,6,7,7,8,8,9,10,10,11,11,12,21,21,21,22,23,23,23,23,23,13,14,14,15,16,16,17,18,18,19,20,20,24,25,25,25,26,26,27,27,27,24,24,28,28,28,26,29,29,29]]
+            #lms[[1,3,4,6,7,9,10,12,13,15,18,20,23,25,38,40,44,46]] += lms[[2,2,5,5,8,8,11,11,14,16,19,21,24,26,39,39,45,45]]
+            #lms[[3,4,6,7,9,10,12,13]] += lms[[5,5,8,8,11,11,14,14]]
+            #lms[[1,15,18,20,23,25,38,40,44,46]] /= 2.0
+            #lms[[3,4,6,7,9,10,12,13]] /= 3.0
+            part_avg = np.mean(np.partition(lms[:,2],3)[0:3])
+            if part_avg < 0.65:
+                avg_conf = part_avg
         return (avg_conf, np.array(lms))
 
     def estimate_depth(self, face_info):
-        lms = np.concatenate((face_info.lms, np.array([[face_info.eye_state[0][1], face_info.eye_state[0][2], face_info.eye_state[0][3]], [face_info.eye_state[1][1], face_info.eye_state[1][2], face_info.eye_state[1][3]]], np.float)), 0)
+        lms = np.concatenate((face_info.lms, np.array([[face_info.eye_state[0][1], face_info.eye_state[0][2], face_info.eye_state[0][3]], [face_info.eye_state[1][1], face_info.eye_state[1][2], face_info.eye_state[1][3]]], np.float32)), 0)
 
         image_pts = np.array(lms)[face_info.contour_pts, 0:2]
 
         success = False
-        if not face_info.rotation is None:
+        if face_info.rotation is not None:
             success, face_info.rotation, face_info.translation = cv2.solvePnP(face_info.contour, image_pts, self.camera, self.dist_coeffs, useExtrinsicGuess=True, rvec=np.transpose(face_info.rotation), tvec=np.transpose(face_info.translation), flags=cv2.SOLVEPNP_ITERATIVE)
         else:
             rvec = np.array([0, 0, 0], np.float32)
@@ -691,54 +783,65 @@ class Tracker():
             face_info.translation = np.array([0.0, 0.0, 0.0], np.float32)
             return False, np.zeros(4), np.zeros(3), 99999., pts_3d, lms
         else:
-            face_info.rotation = np.transpose(face_info.rotation)
+            # OpenCV returns (3,1); flatten so [0]/[1]/[2] are scalars for UDP packing
-            face_info.translation = np.transpose(face_info.translation)
+            # and 1D .dot() math does not broadcast to (1,3).
+            face_info.rotation = np.asarray(face_info.rotation, dtype=np.float32).reshape(3)
+            face_info.translation = np.asarray(face_info.translation, dtype=np.float32).reshape(3)
 
         rmat, _ = cv2.Rodrigues(rotation)
         inverse_rotation = np.linalg.inv(rmat)
-        pnp_error = 0.0
+        t_reference = face_info.face_3d.dot(rmat.transpose())
-        for i, pt in enumerate(face_info.face_3d):
+        t_reference = t_reference + face_info.translation
-            if i == 68:
+        t_reference = t_reference.dot(self.camera.transpose())
+        t_depth = t_reference[:, 2]
+        t_depth[t_depth == 0] = 0.000001
+        t_depth_e = np.expand_dims(t_depth[:],1)
+        t_reference = t_reference[:] / t_depth_e
+        pts_3d[0:66] = np.stack([lms[0:66,0], lms[0:66,1], np.ones((66,))], 1) * t_depth_e[0:66]
+        pts_3d[0:66] = (pts_3d[0:66].dot(self.inverse_camera.transpose()) - face_info.translation).dot(inverse_rotation.transpose())
+        pnp_error = np.power(lms[0:17,0:2] - t_reference[0:17,0:2], 2).sum()
+        pnp_error += np.power(lms[30,0:2] - t_reference[30,0:2], 2).sum()
+        if np.isnan(pnp_error):
+            pnp_error = 9999999.
+        for i, pt in enumerate(face_info.face_3d[66:70]):
+            if i == 2:
                 # Right eyeball
                 # Eyeballs have an average diameter of 12.5mm and and the distance between eye corners is 30-35mm, so a conversion factor of 0.385 can be applied
                 eye_center = (pts_3d[36] + pts_3d[39]) / 2.0
                 d_corner = np.linalg.norm(pts_3d[36] - pts_3d[39])
                 depth = 0.385 * d_corner
                 pt_3d = np.array([eye_center[0], eye_center[1], eye_center[2] - depth])
-                pts_3d[i] = pt_3d
+                pts_3d[68] = pt_3d
                 continue
-            if i == 69:
+            if i == 3:
                 # Left eyeball
                 eye_center = (pts_3d[42] + pts_3d[45]) / 2.0
                 d_corner = np.linalg.norm(pts_3d[42] - pts_3d[45])
                 depth = 0.385 * d_corner
                 pt_3d = np.array([eye_center[0], eye_center[1], eye_center[2] - depth])
-                pts_3d[i] = pt_3d
+                pts_3d[69] = pt_3d
                 continue
-            if i == 66:
+            if i == 0:
-                d1 = np.linalg.norm(lms[i,0:2] - lms[36,0:2])
+                d1 = np.linalg.norm(lms[66,0:2] - lms[36,0:2])
-                d2 = np.linalg.norm(lms[i,0:2] - lms[39,0:2])
+                d2 = np.linalg.norm(lms[66,0:2] - lms[39,0:2])
                 d = d1 + d2
                 pt = (pts_3d[36] * d1 + pts_3d[39] * d2) / d
-            if i == 67:
+            if i == 1:
-                d1 = np.linalg.norm(lms[i,0:2] - lms[42,0:2])
+                d1 = np.linalg.norm(lms[67,0:2] - lms[42,0:2])
-                d2 = np.linalg.norm(lms[i,0:2] - lms[45,0:2])
+                d2 = np.linalg.norm(lms[67,0:2] - lms[45,0:2])
                 d = d1 + d2
                 pt = (pts_3d[42] * d1 + pts_3d[45] * d2) / d
-            reference = rmat.dot(pt)
+            if i < 2:
-            reference = reference + face_info.translation
+                reference = rmat.dot(pt)
-            reference = self.camera.dot(reference)
+                reference = reference + face_info.translation
-            depth = reference[2]
+                reference = self.camera.dot(reference)
-            if i < 17 or i == 30:
+                depth = float(reference[2])
-                reference = reference / depth
+                pt_3d = np.array([lms[66+i][0] * depth, lms[66+i][1] * depth, depth], np.float32)
-                e1 = lms[i][0] - reference[0]
+                pt_3d = self.inverse_camera.dot(pt_3d)
-                e2 = lms[i][1] - reference[1]
+                pt_3d = pt_3d - face_info.translation
-                pnp_error += e1*e1 + e2*e2
+                pt_3d = inverse_rotation.dot(pt_3d)
-            pt_3d = np.array([lms[i][0] * depth, lms[i][1] * depth, depth], np.float32)
+                pts_3d[66+i,:] = pt_3d[:]
-            pt_3d = self.inverse_camera.dot(pt_3d)
+        pts_3d[np.isnan(pts_3d).any(axis=1)] = np.array([0.,0.,0.], dtype=np.float32)
-            pt_3d = pt_3d - face_info.translation
-            pt_3d = inverse_rotation.dot(pt_3d)
-            pts_3d[i,:] = pt_3d[:]
 
         pnp_error = np.sqrt(pnp_error / (2.0 * image_pts.shape[0]))
         if pnp_error > 300:
@@ -761,9 +864,7 @@ class Tracker():
     def preprocess(self, im, crop):
         x1, y1, x2, y2 = crop
         im = np.float32(im[y1:y2, x1:x2,::-1]) # Crop and BGR to RGB
-        im = cv2.resize(im, (self.res_i, self.res_i), interpolation=cv2.INTER_LINEAR) / self.std - self.mean
+        im = cv2.resize(im, (self.res_i, self.res_i), interpolation=cv2.INTER_LINEAR) * self.std_res + self.mean_res
-        #im = cv2.resize(im, (224, 224), interpolation=cv2.INTER_LINEAR) / 255.0
-        #im = (im - mean) / std
         im = np.expand_dims(im, 0)
         im = np.transpose(im, (0,3,1,2))
         return im
@@ -804,7 +905,7 @@ class Tracker():
                 full_frame[0:32, 0:32] = im
             else:
                 full_frame[0:32, 32:64] = im
-        im = im.astype(np.float32)[:,:,::-1] / self.std - self.mean
+        im = im.astype(np.float32)[:,:,::-1] * self.std_32 + self.mean_32
         im = np.expand_dims(im, 0)
         im = np.transpose(im, (0,3,2,1))
         return im, x1, y1, scale, reference, a
@@ -821,11 +922,11 @@ class Tracker():
         x1, y1 = clamp_to_im(center - radius, h, w)
         x2, y2 = clamp_to_im(center + radius + 1, h, w)
         offset = np.array((x1, y1))
-        lms = (lms[:, 0:2] - offset).astype(np.int)
+        lms = (lms[:, 0:2] - offset).astype(int)
         frame = frame[y1:y2, x1:x2]
         return frame, lms, offset
 
-    def get_eye_state(self, frame, lms, single=False):
+    def get_eye_state(self, frame, lms):
         if self.no_gaze:
             return [(1.0, 0.0, 0.0, 0.0), (1.0, 0.0, 0.0, 0.0)]
         lms = np.array(lms)
@@ -841,10 +942,7 @@ class Tracker():
             return [(1.0, 0.0, 0.0, 0.0), (1.0, 0.0, 0.0, 0.0)]
         both_eyes = np.concatenate((right_eye, left_eye))
         results = None
-        if single:
+        results = self.gaze_model.run([], {self.input_name: both_eyes})
-            results = self.gaze_model_single.run([], {self.input_name: both_eyes})
-        else:
-            results = self.gaze_model.run([], {self.input_name: both_eyes})
         open = [0, 0]
         open[0] = 1#results[1][0].argmax()
         open[1] = 1#results[1][1].argmax()
@@ -888,9 +986,15 @@ class Tracker():
             eye_y = eye_y + offset[1]
             eye_state.append([open[i], eye_y, eye_x, conf])
 
+        eye_state = np.array(eye_state)
+        eye_state[np.isnan(eye_state).any(axis=1)] = np.array([1.,0.,0.,0.], dtype=np.float32)
         return eye_state
 
     def assign_face_info(self, results):
+        if self.max_faces == 1 and len(results) == 1:
+            conf, (lms, eye_state), conf_adjust = results[0]
+            self.face_info[0].update((conf - conf_adjust, (lms, eye_state)), np.array(lms)[:, 0:2].mean(0), self.frame_count)
+            return
         result_coords = []
         adjusted_results = []
         for conf, (lms, eye_state), conf_adjust in results:
@@ -945,11 +1049,13 @@ class Tracker():
         self.wait_count += 1
         if self.detected == 0:
             start_fd = time.perf_counter()
-            if self.use_retinaface > 0:
+            if self.use_retinaface > 0 or self.try_hard:
                 retinaface_detections = self.retinaface.detect_retina(frame)
                 new_faces.extend(retinaface_detections)
-            else:
+            if self.use_retinaface == 0 or self.try_hard:
                 new_faces.extend(self.detect_faces(frame))
+            if self.try_hard:
+                new_faces.extend([(0, 0, self.width, self.height)])
             duration_fd = 1000 * (time.perf_counter() - start_fd)
             self.wait_count = 0
         elif self.detected < self.max_faces:
@@ -1003,7 +1109,10 @@ class Tracker():
             output = self.session.run([], {self.input_name: crops[0]})[0]
             conf, lms = self.landmarks(output[0], crop_info[0])
             if conf > self.threshold:
-                eye_state = self.get_eye_state(frame, lms)
+                try:
+                    eye_state = self.get_eye_state(frame, lms)
+                except:
+                    eye_state = [(1.0, 0.0, 0.0, 0.0), (1.0, 0.0, 0.0, 0.0)]
                 outputs[crop_info[0]] = (conf, (lms, eye_state), 0)
         else:
             started = 0
@@ -1065,8 +1174,8 @@ class Tracker():
                 face_info.success, face_info.quaternion, face_info.euler, face_info.pnp_error, face_info.pts_3d, face_info.lms = self.estimate_depth(face_info)
                 face_info.adjust_3d()
                 lms = face_info.lms[:, 0:2]
-                x1, y1 = tuple(lms.min(0))
+                x1, y1 = tuple(lms[0:66].min(0))
-                x2, y2 = tuple(lms.max(0))
+                x2, y2 = tuple(lms[0:66].max(0))
                 bbox = (y1, x1, y2 - y1, x2 - x1)
                 face_info.bbox = bbox
                 detected.append(bbox)
@@ -1092,10 +1201,12 @@ class Tracker():
                         h += 2 * h * self.bbox_growth
                         faces.append((x,y,w,h))
                     self.faces = faces
+        self.faces = [x for x in self.faces if not np.isnan(np.array(x)).any()]
+        self.detected = len(self.faces)
 
         duration = (time.perf_counter() - start) * 1000
         if not self.silent:
-            print(f"Took {duration:.2f}ms (detect: {duration_fd:.2f}ms, crop: {duration_pp:.2f}, track: {duration_model:.2f}ms, 3D points: {duration_pnp:.2f}ms)")
+            print(f"Took {duration:.2f}ms (detect: {duration_fd:.2f}ms, crop: {duration_pp:.2f}ms, track: {duration_model:.2f}ms, 3D points: {duration_pnp:.2f}ms)")
 
         results = sorted(results, key=lambda x: x.id)