[Paper Review] Unsupervised Lifelong Person Re-identification via Contrastive Rehearsal

Hao Chen, Benoit Lagadec+1 more|arXiv (Cornell University)|Mar 12, 2022

Video Surveillance and Tracking Methods5 citations

TL;DR

This paper proposes Unsupervised Contrastive Rehearsal (UCR), a novel method for lifelong person re-identification that enables continual unsupervised domain adaptation without catastrophic forgetting. By contrastively rehearsing a small number of stored old samples and enforcing image-to-image similarity consistency between old and new models, UCR achieves strong generalization on both seen and unseen domains, outperforming prior methods in both zero-shot and in-domain generalization settings.

ABSTRACT

Existing unsupervised person re-identification (ReID) methods focus on adapting a model trained on a source domain to a fixed target domain. However, an adapted ReID model usually only works well on a certain target domain, but can hardly memorize the source domain knowledge and generalize to upcoming unseen data. In this paper, we propose unsupervised lifelong person ReID, which focuses on continuously conducting unsupervised domain adaptation on new domains without forgetting the knowledge learnt from old domains. To tackle unsupervised lifelong ReID, we conduct a contrastive rehearsal on a small number of stored old samples while sequentially adapting to new domains. We further set an image-to-image similarity constraint between old and new models to regularize the model updates in a way that suits old knowledge. We sequentially train our model on several large-scale datasets in an unsupervised manner and test it on all seen domains as well as several unseen domains to validate the generalizability of our method. Our proposed unsupervised lifelong method achieves strong generalizability, which significantly outperforms previous lifelong methods on both seen and unseen domains. Code will be made available at https://github.com/chenhao2345/UCR.

Motivation & Objective

• To address the lack of continual learning in unsupervised person ReID, where models typically forget old knowledge when adapting to new domains.
• To develop a practical lifelong ReID framework that learns incrementally from new, unlabeled data without human annotation.
• To mitigate catastrophic forgetting during unsupervised domain adaptation by preserving old knowledge through contrastive rehearsal.
• To improve generalization to both seen and unseen domains through a unified, scalable training procedure.

Proposed method

• A contrastive rehearsal mechanism stores a small number of representative images and their cluster prototypes from previous domains in long-term memory buffers.
• Pseudo-labels are generated for old domain samples, and contrastive learning is applied to ensure that old samples remain closer to their pseudo-positive pairs than to pseudo-negatives across domain shifts.
• A similarity constraint is enforced between the image-to-image representations of old samples as computed by the frozen old model and the current model, regularizing the update process to preserve old knowledge.
• The method uses a dual loss: a rehearsal loss on stored old samples and a similarity consistency loss between old and new model predictions.
• The model is trained sequentially on multiple large-scale datasets in an unsupervised manner, simulating real-world incremental data arrival.
• Memory update selects the K nearest images to each cluster prototype to reduce pseudo-label noise and improve reliability.

Experimental results

Research questions

• RQ1 Can unsupervised lifelong person ReID be effectively achieved without human annotation on new data?
• RQ2 How can catastrophic forgetting be mitigated during continual unsupervised domain adaptation in ReID?
• RQ3 What role does contrastive rehearsal of old samples play in preserving knowledge across domains?
• RQ4 How does enforcing representation consistency between old and new models improve generalization?
• RQ5 Does the method generalize to unseen domains as well as seen ones?

Key findings

• UCR achieves 59.3% mAP and 75.9% R1 on seen domains, significantly outperforming prior methods like AKA(SL) and C$^2$L(SL).
• On unseen domains, UCR achieves 58.3% mAP and 55.9% R1, demonstrating strong zero-shot generalization capability.
• Increasing the memory size per identity (K$_{mem}$) improves performance, with K$_{mem}$=8 yielding the best results across all metrics.
• The method is robust to training order, maintaining strong performance even when training on the largest domain (MSMT17) first.
• Selecting the nearest images to cluster prototypes for memory storage yields better performance than farthest or random selection.
• Sensitivity analysis shows optimal hyperparameters: τ$_s$=0.2 and λ$_{sim}$=20 for the similarity constraint.