Contrast-Reconstruction Representation Learning for Self-Supervised Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Skeleton-based action recognition is crucial for applications like surveillance and human-machine interaction.
• Supervised methods require extensive labeled data, which is often costly and impractical.
• Unsupervised approaches are needed to overcome the limitations of labeled data dependency.

Purpose of the Study:

• To propose a novel unsupervised method for skeleton-based action recognition.
• To develop a network that simultaneously captures posture and motion dynamics.
• To reduce the dependency on large-scale labeled datasets.

Main Methods:

• Introduced the Contrast-Reconstruction Representation Learning network (CRRL).
• CRRL comprises Sequence Reconstructor (SER) for posture learning and Contrastive Motion Learner (CML) for motion dynamics.
• An Information Fuser (INF) module integrates SER and CML using knowledge distillation for posture-motion coupling.

Main Results:

• The proposed CRRL method achieved state-of-the-art performance on multiple benchmarks (NTU RGB+D 60/120, PKU-MMD, CMU, NW-UCLA).
• Demonstrated effective capture of both static postures and dynamic motion patterns.
• Outperformed existing unsupervised and supervised approaches in skeleton-based action recognition.

Conclusions:

• The CRRL network offers a promising unsupervised solution for skeleton-based action recognition.
• The fusion strategy effectively couples posture and motion representations.
• This method significantly advances unsupervised learning in human action analysis.