yjxiong/tsn-pytorch

Implements TSN for video action recognition with support for multiple modalities—RGB frames, optical flow, and RGB-diff—enabling multi-stream fusion for improved temporal understanding. Uses sparse segment sampling across video clips rather than dense frame processing, with configurable CNN backbones (BNInception, ResNet) and training on standard benchmarks like UCF101. Provides complete training and testing pipelines that maintain compatibility with the original Caffe implementation while leveraging PyTorch's flexibility for research iteration.

1,085 stars. No commits in the last 6 months.