mediapipe_pytorch

PyTorch implementation of Google's Mediapipe model. Iris Landmark model | Face Mesh Model

Face Mesh Model

facial_landmark folder contains the PyTorch implementation of paper Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs (https://arxiv.org/pdf/1907.06724.pdf)

• For inference

  cd facial_landmark
  !python inference.py

Iris Landmark Model

iris folder contains the PyTorch implementation of paper Real-time Pupil Tracking from Monocular Video for Digital Puppetry (https://arxiv.org/pdf/2006.11341)

• For inference

  cd iris
  !python inference.py

Conversion Issues

* TFLite uses slightly different padding compared to PyTorch.
* Instead of using the padding parameter in the conv layer applying padding manually.
* Change the padding value.
    * Misleading results
        * nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, stride=2, padding=1, bias=True)
    * Correction
        * nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, stride=2, padding=0, bias=True)
            * x = nn.ReflectionPad2d((1, 0, 1, 0))(x) # Apply padding before convolution.