Reproduce “Actionlet-Dependent Contrastive Learning for Unsupervised Skeleton-Based Action Recognition” on OpenResearch

About this paper

Lilang Lin, Jiahang Zhang, Jiaying Liu

Actionlet-Dependent Contrastive Learning (ActCLR) improves unsupervised skeleton-based action recognition by identifying motion-relevant joints and time intervals without action labels. It uses these actionlets to preserve semantic motion during augmentation, apply stronger transformations to less informative regions, and focus feature pooling on discriminative body parts, achieving up to 88.8% linear-evaluation accuracy on NTU RGB+D 60 and consistent gains in transfer and action-segmentation settings.

Key insights

• Separating action-relevant and less-relevant regions makes aggressive augmentation more useful: transformations can improve robustness without destroying the motion cues needed to identify an action.
• A dataset-average skeleton can serve as a label-free static reference for gradient-based localization, although the assumption is less suitable for whole-body or poorly aligned motions.
• Actionlet-aware pooling improves representation quality when used in the offline branch; applying the same focus in the online branch can interfere with learning from transformed non-actionlet regions.

Summary by alphaXiv.