- Enable ORT_ENABLE_BASIC graph optimization instead of fully disabling it, yielding CPU fallback performance improvements without crashing the GPU on FusedConv.
- Set cv2.CAP_PROP_BUFFERSIZE to 1 on Linux VideoCapture to eliminate webcam tracking lag/latency.
1. Fix shape broadcasting crash in Numpy 2.0 with .flatten()
2. Fix cv2.solvePnP in OpenCV 5.0 by setting useExtrinsicGuess=False on the first frame
3. Switch default models to _gpu.onnx to avoid FusedConv Clip crashes on CUDAExecutionProvider
I based changed the numbers based on lengths I observed with `print(len(f.lms))` and `print(len(f.pts_e3))`. Once I changed the numbers to their new values, the header became the same length as the data, which makes me think I reconstructed the intent correctly. However, since the readme says the training data had 66 key points, it raises questions about what the extra 2 and 4 keypoints respectively represent.
This fixes the following error on exit when OpenCV is built without
GUI support:
Traceback (most recent call last):
File "facetracker.py", line 450, in <module>
cv2.error: OpenCV(4.9.0) /build/source/modules/highgui/src/window.cpp:1266: error: (-2:Unspecified error) The function is not implemented. Rebuild the library with Windows, GTK+ 2.x or Cocoa support. If you are on Ubuntu or Debian, install libgtk2.0-dev and pkg-config, then re-run cmake or configure script in function 'cvDestroyAllWindows'
[254225] Failed to execute script 'facetracker' due to unhandled exception!
When building with the latest numpy version (v1.24.0), OpenSeeFace crashes with "AttributeError: module 'numpy' has no attribute 'float'". This is because the attribute 'float' is now deprecated in numpy v1.24. I have changed the requirements to the earlier version v1.23.0. This should probably only be a temporary fix, as I'm sure there is a good reason that float was depreciated. I would also recommend changing the requirements to a single file with a static version for each requirement that is known to work. This will help ensure the stability of the program. I may do this in another fork when I have more time if you want.
When building with the latest numpy version (v1.24.0), OpenSeeFace crashes with "AttributeError: module 'numpy' has no attribute 'float'". This is because the attribute 'float' is now deprecated in numpy v1.24. I have changed the requirements to the earlier version v1.23.0. This should probably only be a temporary fix, as I'm sure there is a good reason that float was depreciated. I would also recommend changing the requirements to a single file with a static version for each requirement that is known to work. This will help ensure the stability of the program. I may do this in another fork when I have more time if you want.
