Knowledge-Preserving continual person re-identification using Graph Attention Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (ReID) is crucial for intelligent security but current deep models require retraining for new scenarios, leading to catastrophic forgetting.
• Catastrophic forgetting hinders the adaptability of deep learning models when encountering new data domains.

Purpose of the Study:

• To propose a Continual person re-identification model via a Knowledge-Preserving (CKP) mechanism.
• To enable deep learning models to accumulate knowledge from continuously changing scenarios without forgetting previous learning.

Main Methods:

• Developed a Continual person re-identification model (CKP) that accumulates knowledge from new domains.
• Utilized a graph attention network, inspired by human cognition, to update knowledge as scenarios change.
• Employed accumulated knowledge to guide the learning process on new image samples.

Main Results:

• CKP demonstrated significant advantages in preventing catastrophic forgetting across various person re-identification benchmark datasets.
• The model achieved superior performance on both seen and never-seen domains compared to fine-tuning, continual learning, and joint training methods.
• CKP outperformed the best comparative model by up to 1.02% in mAP and Rank1 metrics on unseen domains.

Conclusions:

• The proposed CKP mechanism effectively addresses catastrophic forgetting in continual person re-identification.
• CKP exhibits strong generalization ability, performing well on datasets not encountered during training.
• The approach offers a promising solution for adaptive and robust intelligent security systems.