Distribution-Aware Knowledge Aligning and Prototyping for Non-Exemplar Lifelong Person Re-Identification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Lifelong person re-identification (LReID) is challenged by catastrophic forgetting in non-stationary data streams.
• Existing methods face limitations in data privacy (exemplar-based) or acquisition capacity (knowledge distillation-based).
• Prototypes are under-investigated in LReID, with prior work neglecting crucial instance-level diversity and distribution information.

Purpose of the Study:

• To introduce a novel framework, Distribution-aware Knowledge Aligning and Prototyping (DKP++), for effective lifelong person re-identification.
• To address the limitations of existing LReID methods by better balancing knowledge retention and acquisition using prototypes.
• To overcome challenges posed by domain shifts and feature gaps in leveraging fine-grained prototype information.

Main Methods:

• Instance-level Distribution Modeling network to capture local diversity within each instance.
• Distribution-oriented Prototype Generation algorithm to create identity-level distribution prototypes.
• Prototype-based Knowledge Transfer module for distilling knowledge from historical prototypes.
• Privacy-friendly Distribution Aligning module with feature-level constraints to mitigate domain shifts.

Main Results:

• DKP++ effectively models and leverages instance-level diversity and identity-level distributions as prototypes.
• The Distribution Aligning module successfully mitigates domain shifts, enhancing historical prototype utilization.
• The proposed method achieves a superior balance between plasticity and stability in lifelong learning scenarios.

Conclusions:

• DKP++ significantly outperforms state-of-the-art LReID methods by a large margin.
• The framework demonstrates the potential of distribution-aware prototypes for robust lifelong person re-identification.
• DKP++ offers a promising direction for addressing catastrophic forgetting in dynamic environments.