kenziyuliu/MS-G3D

Implements a multi-stream graph convolutional architecture that separately models joint and bone sequences, with support for two-stream fusion to improve skeleton-based action recognition. The approach uses spatial-temporal graph convolutions with configurable topology, trained via mixed-precision (NVIDIA Apex) on NTU RGB+D and Kinetics datasets. Includes pretrained models and joint-bone ensemble capabilities for reproducible results across multiple evaluation protocols.

455 stars. No commits in the last 6 months.