Bi-C²R: Bidirectional Continual Compatible Representation for Re-Indexing Free Lifelong Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Lifelong person Re-IDentification (L-ReID) continuously updates models with new data but faces catastrophic forgetting.
• Existing L-ReID methods require costly re-indexing of historical gallery images after each model update.
• Data privacy concerns and high computational costs hinder the re-indexing of large-scale gallery datasets.

Purpose of the Study:

• Introduce a new task, Re-index Free Lifelong person Re-IDentification (RFL-ReID), to address the limitations of traditional L-ReID.
• Develop a method that performs lifelong person re-identification without re-indexing historical gallery data.
• Ensure compatibility between features extracted by updated and previous models.

Main Methods:

• Propose the Bidirectional Continuous Compatible Representation (Bi-C²R) framework.
• Design a bidirectional compatible transfer network to update old gallery features to the new feature space.
• Incorporate bidirectional compatible distillation and anti-forgetting distillation modules to balance new and old knowledge.
• Utilize a feature-level exponential moving average strategy to bridge knowledge gaps across diverse data domains.

Main Results:

• The Bi-C²R framework effectively updates gallery features without re-indexing.
• The proposed method balances compatibility between new and old knowledge in dual feature spaces.
• Experimental results demonstrate leading performance on both RFL-ReID and traditional L-ReID tasks.
• Theoretical analysis and extensive experiments validate the method's effectiveness on multiple benchmarks.

Conclusions:

• The Bi-C²R framework offers an efficient and compatible solution for lifelong person re-identification without re-indexing.
• RFL-ReID presents a more practical and scalable approach to lifelong person re-identification.
• The method successfully mitigates catastrophic forgetting and ensures feature compatibility in continuous learning scenarios.