Open Turlan opened 1 year ago
I have also noticed this issue. In fact, when running demo.py, there is no preprocessing operation performed on the input data. However, it is done during training, which can introduce errors here. In practice, demo.py cannot be used to predict videos.
The preprocessing is not aligned with your training in your demo code. For the data preprocessing in training, the face images are first cropped to squares. It seems that u just ignore this and resize the image without keeping the aspect ratio in the demo.py.