A Self-Supervised Gait Encoding Approach With Locality-Awareness for 3D Skeleton Based Person Re-Identification.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Person re-identification (Re-ID) using gait from 3D skeletons is an emerging field.
• Current methods often rely on hand-crafted features or supervised learning, limiting their adaptability.
• Unlabeled 3D skeleton data offers a valuable resource for gait representation learning.

Purpose of the Study:

• To propose a novel self-supervised gait encoding approach for person Re-ID using unlabeled 3D skeleton sequences.
• To leverage the inherent locality of motion in 3D skeleton data for improved representation learning.
• To introduce Constrastive Attention-based Gait Encodings (CAGEs) for effective gait representation.

Main Methods:

• Developed a self-supervised learning framework by reconstructing skeleton sequences reversely.
• Introduced a locality-aware attention mechanism to capture intra-sequence motion correlations.
• Implemented a locality-aware contrastive learning scheme to preserve inter-sequence temporal correlations.
• Designed Constrastive Attention-based Gait Encodings (CAGEs) using learned context vectors.

Main Results:

• The self-supervised approach significantly outperforms existing skeleton-based person Re-ID methods.
• Achieved 15-40 percent improvement in Rank-1 accuracy compared to skeleton-based counterparts.
• Demonstrated superior performance even against multi-modal methods incorporating RGB or depth data.

Conclusions:

• Self-supervised learning effectively extracts gait representations from unlabeled 3D skeleton data for person Re-ID.
• The proposed locality-aware mechanisms enhance the learning of gait features by preserving motion continuity.
• CAGEs offer a powerful and efficient method for gait-based person Re-ID, advancing the state-of-the-art.