Worse performance of released pytorch models

[XChenZ]

Hello emilianavt, I'm trying to utilize the provided pytorch models (model.py) and weights to detect facial landmarks. However, I found its performance is worse than onnx version. An example is shown below (lm_model3 for both).

Do you have any suggestion? In my practice, I modify the model.py according to this issue, and I adapt the tracker.py by replacing the onnx models (self.session and self.detection) with pytorch models (OpenSeeFaceLandmarks and OpenSeeFaceDetect). I follow the image preprocessing codes and postprocessing codes in the original tracker.py. I'm wondering if you could provide a demo tracking code for pytorch models. Any advice is appreciated!