temporal-shift-module

Why we included this project

Video understanding usually means paying a steep compute bill, but TSM shows how to get most of the benefit for a fraction of the cost. Its temporal shift module moves feature maps across time so a plain 2D CNN can model motion between frames without 3D convolutions or optical flow, which is what makes the approach attractive on edge devices. The repo has PyTorch training code, pre-trained Kinetics weights, and the setup needed to reproduce the published numbers or fine-tune for your own action recognition. There is even an online demo running real-time hand gesture recognition on a Jetson Nano at 8 watts, so you can see the method work on actual hardware rather than just in a paper.