Lifelong person re-identification via dynamically knowledge adaptation and retention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (ReID) models need continuous learning for real-world applications.
• Lifelong ReID faces catastrophic forgetting, hindering performance on new data.
• Existing methods struggle to balance performance across old and new domains.

Purpose of the Study:

• To propose a Dynamic Knowledge Adaptation and Retention framework for lifelong ReID.
• To address the limitations of existing methods in balancing performance between old and new domains.
• To mitigate catastrophic forgetting while enabling adaptation to new data.

Main Methods:

• A shared backbone network extracts features, followed by a Dynamic Adaptation module for instance normalization and adaptive kernel generation.
• An Adaptability Retention strategy restricts model updates using a frozen model as a knowledge anchor.
• A constraint loss maintains parameter consistency between the knowledge anchor and the trainable model.

Main Results:

• The proposed framework achieves dynamic learning and adaptive feature optimization.
• It effectively balances knowledge retention and adaptation to new domains.
• Demonstrated outstanding performance on four mainstream ReID datasets in a lifelong setting.

Conclusions:

• The Dynamic Knowledge Adaptation and Retention framework successfully mitigates catastrophic forgetting in lifelong ReID.
• The method achieves superior performance with an average Rank-1 accuracy of 65.7% and mAP of 55.4%.
• This approach offers a robust solution for continuous learning in person re-identification systems.