Diverse Representations Embedding for Lifelong Person Re-Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Lifelong person re-identification (LReID) continuously learns from sequential data for cross-camera matching.
• Key challenges include balancing knowledge preservation with new information acquisition and limited representation capacity.
• Conventional CNN-based methods often fail to capture diverse instance perspectives, impacting performance.

Purpose of the Study:

• To propose a Diverse Representation Embedding (DRE) framework for effective LReID.
• To address the limitations of conventional methods in preserving knowledge and adapting to new data.
• To enhance the representation capacity and performance of LReID models.

Main Methods:

• Developed a Transformer-based backbone utilizing maximum embedding (ME) and multiple class tokens for overlapping instance representations.
• Designed an Adaptive Constraint Module (ACM) for integrating and discriminating overlapping representations to create diverse embeddings.
• Implemented Knowledge Update (KU) and Knowledge Preservation (KP) strategies within adjustment and learner models.

Main Results:

• The DRE framework demonstrated superior performance compared to state-of-the-art methods on large-scale, occluded, and holistic datasets.
• Significant performance gains were achieved, validating the effectiveness of the proposed approach.
• The method successfully balances the preservation of old knowledge with adaptation to new information.

Conclusions:

• The proposed DRE framework effectively tackles the challenges in LReID by enhancing representation diversity and managing knowledge.
• The Transformer-based backbone, ACM, and KU/KP strategies contribute to improved LReID performance.
• This work offers a promising direction for continuous learning in person re-identification systems.