This is a Pytorch implementation of Unsupervised Learning of Optical Flow via Brightness Constancy and Motion Smoothness paper. The Original paper uses FlownetS as a backbone supervised model. In this project the original paper pipeline have been implemented, and other supervised models are to be added; Light Flownet and PWC NET
Since state of the art of supervised optical flow estimation changes frequently. You can contribute to this project by adding the supervised model implementation to models.py following the syntaxe of the supervised models already implemented. Namely you need to make sure that your forward pass returns a batch of optical flow estimation of size (N, 2, H, W), or a tuple of batches of different spatial resolutions (N, 2, h1, w1), (N, 2, h2, w2) ...