Reproduce “3D Human Action Representation Learning via Cross-View Consistency Pursuit” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, Wenjun Zhang
An unsupervised cross-view contrastive learning framework, CrosSCLR, improves 3D human action representation by exploiting multi-view skeleton data (joint, motion, bone) to mine high-confidence positive samples and enforce consistency across views. The framework shows enhanced performance in action recognition on NTU-RGB+D datasets and offers superior initialization for supervised fine-tuning.
Key insights
- Leveraging complementary information across multiple skeleton views (joint, motion, bone) significantly improves the quality and discriminative power of unsupervised 3D action representations.
- Expanding the positive set in contrastive learning beyond direct augmentations by mining "hidden" high-confidence positive samples from a memory bank helps form more regular and semantically coherent embedding spaces.
- Enforcing consistency of embedding distributions, not just positive sample labels, across different views is crucial for integrating multi-view knowledge and improving overall representation robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.