Problem
Research questions and friction points this paper is trying to address.
Addresses class imbalance in action recognition
Improves representation learning for tail classes
Enhances generalization in skeleton-based datasets
Balanced Representation Learning for Long-tailed Skeleton-based Action Recognition
🤖 AI Summary
To address the long-tailed distribution problem in skeleton-based action recognition, this paper proposes a decoupled representation learning framework that jointly optimizes class balance in both feature space and classifier space. The method innovatively integrates representation disentanglement with dynamic class reweighting, incorporating prototype alignment regularization and tail-class-aware contrastive learning to effectively mitigate feature shift and classifier bias. Built upon graph convolutional networks (GCNs), it employs a prototype memory bank, dynamic label smoothing, and class-aware feature reweighting. Evaluated on the NTU-60 and NTU-120 long-tailed benchmarks, the approach achieves an overall accuracy improvement of 5.2% and a substantial 12.7% gain in tail-class accuracy, surpassing existing state-of-the-art methods.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Spatial-temporal action exploration strategy
Detached action-aware learning schedule
Skip-modal representation for structural information
🔎 Similar Papers
No similar papers found.
