MS-CLR: Multi-Skeleton Contrastive Learning for Human Action Recognition

TL;DR

This paper introduces MS-CLR, a self-supervised contrastive learning framework that aligns pose representations across multiple skeleton conventions, enhancing generalization and robustness in human action recognition.

Contribution

MS-CLR is the first to explicitly align multiple skeleton conventions in contrastive learning, improving generalization across datasets with diverse joint structures.

Findings

MS-CLR outperforms single-skeleton contrastive methods on NTU datasets.

Multi-skeleton ensemble achieves state-of-the-art results.

The unified representation scheme handles varying skeleton structures effectively.

Abstract

Contrastive learning has gained significant attention in skeleton-based action recognition for its ability to learn robust representations from unlabeled data. However, existing methods rely on a single skeleton convention, which limits their ability to generalize across datasets with diverse joint structures and anatomical coverage. We propose Multi-Skeleton Contrastive Learning (MS-CLR), a general self-supervised framework that aligns pose representations across multiple skeleton conventions extracted from the same sequence. This encourages the model to learn structural invariances and capture diverse anatomical cues, resulting in more expressive and generalizable features. To support this, we adapt the ST-GCN architecture to handle skeletons with varying joint layouts and scales through a unified representation scheme. Experiments on the NTU RGB+D 60 and 120 datasets demonstrate that MS-CLR consistently improves performance over strong single-skeleton contrastive learning baselines. A multi-skeleton ensemble further boosts performance, setting new state-of-the-art results on both datasets.