Compositional Lipschitz Bounds and Robustness for Multi-Stream Skeleton-Based Action Recognition

The strongest results in skeleton-based human action recognition are almost always produced by multi-stream ensembles, which average the softmax outputs of parallel joint, bone and velocity branches; despite this, the operator that such an ensemble defines has never been given aprovable stability guarantee. We provide a self-contained compositional Lipschitz analysis of this operator, viewing the multi-stream recognizer as a convex aggregation of parallel branches in which each branch composes a hand-designed stream generator, a deep graph--temporal encoder and a softmax classifier. The central result is that the Lipschitz constant of the full operator coincides with the weight-averaged branch constant.