Overcoming Dual Incremental Challenges in Continual Person Search via Adapter and Prototype.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Continual person search combines person detection and re-identification (Re-ID), facing domain and class incremental learning challenges.
• Existing methods struggle with adapting to new domains and identity classes without forgetting previous knowledge.

Purpose of the Study:

• To develop a novel framework addressing the dual incremental challenges in continual person search.
• To improve both person detection and Re-ID performance within a continual learning paradigm.

Main Methods:

• Proposed a framework utilizing an adapter-based Swin Transformer backbone.
• Introduced Domain Aware Adapter (DAA) blocks with a Domain Prototype Router (DPR) for domain incremental learning.
• Implemented Virtual Prototype Replay-Online Instance Matching (VPR-OIM) for class incremental Re-ID, incorporating virtual prototype replay.

Main Results:

• The DAA and VPR-OIM components effectively address the dual incremental challenges of continual person search.
• Experimental results show significant improvements in both person detection and Re-ID performance under continual learning settings.
• The proposed method achieves state-of-the-art (SOTA) performance in continual person search tasks.

Conclusions:

• The novel framework successfully mitigates catastrophic forgetting in continual person search.
• The approach demonstrates superior performance in handling both domain and class increments.
• This work offers a robust solution for real-world applications requiring adaptive person search systems.