3D-ResNets-PyTorch

Why we included this project

This codebase handles a specific computer vision problem: figuring out what people are doing in video clips rather than just what objects appear in a frame. It ships complete PyTorch training, fine-tuning, and evaluation scripts for 3D residual networks, so you can go from raw video to a working action-recognition model without building the architecture and data pipeline yourself. Pretrained weights cover several mainstream benchmarks, including Kinetics, Moments in Time, UCF-101, HMDB-51, and ActivityNet, which helps when you have limited labeled footage and would rather adapt an existing model than train from scratch. The code maps directly onto the authors' CVPR and ICCV papers, so reproducing published results and building on them is straightforward.