mmaction2

Why we included this project

mmaction2 is what we reach for when a video task outgrows a single model and becomes a real pipeline problem. Built on PyTorch as part of OpenMMLab, it gathers the field's common building blocks in one place: TSN, TSM, SlowFast and X3D for action recognition, BMN-style temporal localization, spatio-temporal detectors built for AVA, and skeleton-based recognizers like PoseC3D. Instead of stitching separate codebases together, a team can compose these pieces to classify activities, pin down when an action starts and ends, or track who is doing what across frames. The modular design also makes it far more practical to adapt a proven recipe to a custom dataset than to start from zero. The model zoo and solid documentation give researchers and engineers a consistent base to build on.