Separable Spatial-Temporal Residual Graph for Cloth-Changing Group Re-Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Group re-identification (GReID) is vital for video surveillance but struggles with member appearance changes.
• Existing GReID methods fail in long-term surveillance due to cloth-changing behaviors.
• A new task, cloth-changing group re-identification (CCGReID), is proposed to address this limitation.

Purpose of the Study:

• To develop a robust group re-identification method capable of handling cloth-changing members.
• To introduce a novel graph-based approach for modeling spatial and temporal group relationships.
• To advance the field of GReID with improved accuracy and resilience to appearance variations.

Main Methods:

• Proposing the separable spatial-temporal residual graph (SSRG) for CCGReID.
• Constructing spatial member graphs (SMG) for intra-image group features and temporal member graphs (TMG) for inter-image feature propagation.
• Utilizing residual learning for efficient SMG and TMG training and enabling inference-time application.

Main Results:

• SSRG achieves state-of-the-art performance on CCGReID tasks, demonstrating superior accuracy.
• The method exhibits low performance degradation (2.15% on GroupVC) despite cloth changes.
• SSRG generalizes well to standard GReID tasks and surpasses some supervised methods in weakly supervised settings.

Conclusions:

• The proposed SSRG effectively models group relationships and enhances robustness against cloth-changing members in GReID.
• SSRG offers a significant advancement for long-term video surveillance applications.
• The developed datasets and method pave the way for future research in CCGReID.