Learning representations of human actions from skeleton sequences, using hyperbolic geometry and hierarchy-aware learning.
Human motion can be represented as sequences of joint positions. This project explores learning representations of those sequences using hyperbolic geometry and hierarchy-aware learning.
Approach
HypSkeletonCLR-HPL extends Hyperbolic SkeletonCLR with hierarchy-aware pseudo-label learning. The implementation combines clustering, prototype objectives, and pseudo-label feedback for skeleton-based action representation learning on NTU RGB+D data.
Implementation
The repository includes pre-training, linear evaluation, and tools for inspecting embeddings and hierarchy diagnostics. The source code and configuration are available through the GitHub icon above.