GCReID: Generalized continual person re-identification via meta learning and knowledge accumulation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (ReID) models struggle with catastrophic forgetting when adapting to new domains.
• Existing continual learning methods for ReID have limited generalization due to domain distribution differences.

Purpose of the Study:

• To propose a generalized continual person re-identification (GCReID) model for anti-forgetting and generalizable ReID.
• To enhance model generalization by simulating unseen domains and employing meta-learning.

Main Methods:

• The GCReID model simulates unseen domains to increase sample diversity.
• Meta-training and meta-testing strategies are used to boost generalization.
• A graph attention network integrates universal knowledge from seen and simulated domains.

Main Results:

• The proposed GCReID model effectively alleviates catastrophic forgetting in seen domains.
• Experiments on 12 benchmark datasets show significant improvements in generalization to unseen domains.
• GCReID outperforms 6 representative ReID models in extensive comparisons.

Conclusions:

• The GCReID model offers a robust solution for continual learning in person re-identification.
• The approach successfully balances anti-forgetting capabilities with enhanced generalization across diverse domains.