Anti-Forgetting Adaptation for Unsupervised Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation for person re-identification (ReID) struggles with knowledge retention and generalization to new datasets.
• Existing ReID models often forget previously acquired knowledge when adapting to new domains.

Purpose of the Study:

• To propose a novel framework, Dual-level Joint Adaptation and Anti-forgetting (DJAA), for incremental unsupervised domain adaptive person ReID.
• To enable ReID models to adapt to new domains while retaining knowledge from source and previously adapted target domains.

Main Methods:

• The DJAA framework employs prototype and instance-level consistency for adaptation.
• A memory buffer stores representative image samples and cluster prototypes, updated incrementally.
• Regularization techniques enforce image-to-image and image-to-prototype similarity to rehearse past knowledge.

Main Results:

• The proposed DJAA framework significantly enhances anti-forgetting capabilities in unsupervised person ReID.
• Experiments show improved generalization ability to unseen domains after multi-step adaptation.
• The method demonstrates strong backward-compatible performance across all seen domains.

Conclusions:

• The DJAA framework effectively addresses the forgetting problem in incremental unsupervised domain adaptive person ReID.
• DJAA offers a robust solution for maintaining and improving ReID model performance across evolving datasets.
• This approach advances the state-of-the-art in domain adaptive person ReID by ensuring continuous learning and broad applicability.