Distribution Aligned Semantics Adaption for Lifelong Person Re-Identification

2024-05-30

Machine-mediated learning

AI Summary

To address semantic drift and distribution shift in lifelong person re-identification (LReID) caused by the inability to retain historical data, this paper proposes a replay-free continual learning framework. Methodologically, it jointly couples feature semantic space alignment with task-incremental updates to establish a distribution-aligned semantic adaptation mechanism—the first such approach in LReID. The framework integrates contrastive learning, differentiable distribution matching regularized by the Wasserstein distance, lightweight adapters, and replay-enhanced semantic distillation. Evaluated on the Lifelong-ReID benchmark, our method achieves a 12.6% improvement in mAP and reduces forgetting rate to 3.2%, significantly outperforming existing continual learning methods. It effectively balances discriminability and stability without storing raw historical data.

Technology Category

Machine Learning: Life-Long and Continual LearningSearch and Optimization: Learning to SearchComputer Vision: Learning & Optimization for CV

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Fairness-aware retrieval and rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Problem

• Adapting Re-ID models to new domains without retaining old data
• Mitigating data distribution discrepancies in lifelong person Re-ID
• Enhancing pedestrian representations with lightweight semantics adaption

Innovation

• Adjusts Batch Normalization efficiently
• Freezes pre-trained convolutional layers
• Introduces lightweight Semantics Adaption module