Abstract
Unsupervised Person Re-identification (ReID) models typically struggle with "catastrophic forgetting"—the loss of knowledge about previous environments when adapting to new ones. This paper introduces the Dual-level Joint Adaptation and Anti-forgetting (DJAA) framework, a lifelong learning approach for ReID. DJAA employs a hybrid memory buffer and dual-level consistency regularization to retain old knowledge while adapting to new unlabeled domains. Experimental results demonstrate that DJAA significantly outperforms state-of-the-art methods in anti-forgetting, generalization to unseen domains, and backward compatibility.
Introduction
Person Re-identification (ReID) is a critical task in computer vision, involving the matching of individuals across non-overlapping camera views. While Unsupervised Domain Adaptation (UDA) has advanced the field, most models are designed for static environments. In real-world applications, surveillance systems are dynamic, requiring models to learn from a continuous stream of data. Standard adaptation techniques often lead to catastrophic forgetting, where the model's performance on earlier domains degrades as it learns from new data.
Problem Statement
The authors identify three primary challenges in lifelong unsupervised ReID:
- Catastrophic Forgetting: Loss of fine-grained features from previous domains during adaptation.
- Poor Generalization: Overfitting to specific training styles, leading to failure in entirely new environments.
- Backward Incompatibility: The inability of updated models to match new "query" features against archived "gallery" features without costly re-extraction.
Methodology: The DJAA Framework
The DJAA framework addresses these challenges through two integrated modules: the Adaptation Module and the Rehearsal Module.
Adaptation Module
This module facilitates learning from the current unlabeled domain. It utilizes DBSCAN clustering to generate pseudo-labels and applies two contrastive losses:
- Prototype-level Loss ($L_{pa}$): Aligns image features with their respective cluster prototypes (centroids).
- Instance-level Loss ($L_{ia}$): Encourages similarity between "hard positive" pairs to refine identity representations.
Rehearsal Module
To mitigate forgetting, a hybrid memory buffer stores representative images and prototypes from prior domains. The core innovation is Consistency Regularization, which ensures that the similarity relationships between images and prototypes remain stable between the previous (frozen) model and the current (trainable) model. This is implemented at both the Image-to-Prototype ($L_{ps}$) and Image-to-Image ($L_{is}$) levels.
Experimental Evaluation
The framework was evaluated using a suite of "seen" domains for training (Market1501, CUHK-SYSU, MSMT17) and 10 "unseen" domains for testing generalization (e.g., VIPeR, PRID, GRID).
Key Results
DJAA demonstrated superior performance across all metrics:
- Anti-Forgetting: On the Market1501 dataset, DJAA maintained a 65.2% mAP after multiple adaptation steps, compared to only 29.0% for the baseline ICE model.
- Generalization: DJAA achieved an average mAP of 57.1% across 10 unseen domains, surpassing existing lifelong learning methods by over 6%.
- Backward Compatibility: The model maintained stable feature representations, allowing new queries to be effectively matched against old gallery images.
Conclusion
The DJAA framework provides a robust solution for lifelong unsupervised person ReID. By combining prototype-based rehearsal with dual-level consistency, it effectively balances the acquisition of new knowledge with the preservation of old expertise. Future work may focus on reducing memory requirements through generative rehearsal techniques.
Summary updated