Cross-stream contrastive learning for self-supervised skeleton-based action recognition

Ding Li, Yongqiang Tang, Zhizhong Zhang, Wensheng Zhang

Abstract: Highlights•Contrastive skeleton-based recognition methods tend to utilize easy positive pairs.•The defect of single-stream contrast and the lack of necessary feature transformation result in easy positives.•Cross-stream contrastive learning introduces hard positives.•Synthetic positive pairs can be generated by feature-level transformation.