sign-language-recognition

Dataset

WLASL dataset: [https://arxiv.org/abs/1910.11006], [https://github.com/dxli94/WLASL]

Returns videos as torch.FloatTensor of shape (C x T x H x W)

• C - number of channels, 3 (rgb)
• T - number of consecutive frames of video (currently set to constant 50)
• H x W - height x widht of frames, set to 224 x 224 due to RandomCrop

I3D model

[https://arxiv.org/abs/1705.07750], [https://github.com/piergiaj/pytorch-i3d]

TODO

• train conv2d-rnn with optimal parameters
• train pose-rnn with optimal parameters
• add option to load subset of WLASL dataset (WLASL100, WLASL300, .. - described in WLASL paper).
• implement I3D model
• implement pose-rnn model