PWC-Net

Why we included this project

PWC-Net is the reference implementation of a well-known 2018 CVPR paper that fused classic optical flow techniques into a single end-to-end trainable network. Anyone doing motion analysis in video, for tasks like action recognition, frame interpolation, or autonomous driving perception, gets both Caffe and PyTorch code here, with the two implementations closely matching each other's accuracy on benchmarks such as Sintel. The repo also ships pretrained models and the full training and testing pipeline, so it works as a real starting point rather than just a paper reference. One caveat: the code targets Python 2.7 and an older deep learning stack, so it is best treated as a documented baseline for reproducing published results, or as a source of architecture ideas to port into a modern framework.