Human Motion Segmentation via Velocity-Sensitive Dual-Side Auto-Encoder.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Human motion segmentation (HMS) is crucial for analyzing human actions in videos.
• Supervised methods require costly large datasets, while unsupervised methods struggle with temporal correlations.

Purpose of the Study:

• To develop a novel unsupervised framework for human motion segmentation.
• To overcome limitations of existing supervised and unsupervised approaches.

Main Methods:

• Proposed Velocity-Sensitive Dual-Side Auto-Encoder (VSDA) framework.
• Utilized Multi-Neighbor Auto-Encoder (MNA) for temporal feature extraction.
• Implemented Long-Short distance Encoding (LSE) and Decoding (LSD) strategies.
• Incorporated a velocity-sensitive (VS) guidance mechanism.

Main Results:

• VSDA effectively extracts informative temporal features from human motions.
• The LSE/LSD strategy enforces distinct representations for frames at varying temporal distances.
• The VS guidance mechanism further enhances model performance by considering energy variations.

Conclusions:

• The proposed VSDA framework achieves promising performance in human motion segmentation.
• Demonstrated effectiveness across six real-world human motion datasets.
• Offers an effective unsupervised alternative for HMS tasks.