PaddleVideo provides video understanding models and tools for action recognition, temporal action localization, and skeleton-based activity detection. It includes implementations of popular architectures like SlowFast, ST-GCN, TSM, and TSN, along with the project’s PP-TSM and PP-TSMv2 lightweight action recognition models. The toolkit supports video data annotation and covers datasets like Kinetics-400, AVA, ActivityNet, and YouTube-8M.