[Paper Review] Unsupervised Person Re-Identification: A Systematic Survey of Challenges and Solutions

Xiangtan Lin, Pengzhen Ren+4 more|arXiv (Cornell University)|Sep 1, 2021

Video Surveillance and Tracking MethodsComputer Science108 references18 citations

TL;DR

This survey provides a comprehensive analysis of unsupervised person re-identification (Re-ID) methods, systematically addressing four core challenges: lack of identity labels, pseudo-supervision for feature learning, cross-camera domain shift, and domain gaps between datasets. It evaluates state-of-the-art solutions, summarizes performance across benchmarks, and identifies key research gaps and future directions for scalable, generalizable Re-ID models using unlabelled data.

ABSTRACT

Person re-identification (Re-ID) has been a significant research topic in the past decade due to its real-world applications and research significance. While supervised person Re-ID methods achieve superior performance over unsupervised counterparts, they can not scale to large unlabelled datasets and new domains due to the prohibitive labelling cost. Therefore, unsupervised person Re-ID has drawn increasing attention for its potential to address the scalability issue in person Re-ID. Unsupervised person Re-ID is challenging primarily due to lacking identity labels to supervise person feature learning. The corresponding solutions are diverse and complex, with various merits and limitations. Therefore, comprehensive surveys on this topic are essential to summarise challenges and solutions to foster future research. Existing person Re-ID surveys have focused on supervised methods from classifications and applications but lack detailed discussion on how the person Re-ID solutions address the underlying challenges. This survey review recent works on unsupervised person Re-ID from the perspective of challenges and solutions. Specifically, we provide an in-depth analysis of highly influential methods considering the four significant challenges in unsupervised person Re-ID: 1) lacking ground-truth identity labels to supervise person feature learning; 2) learning discriminative person features with pseudo-supervision; 3) learning cross-camera invariant person feature, and 4) the domain shift between datasets. We summarise and analyse evaluation results and provide insights on the effectiveness of the solutions. Finally, we discuss open issues and suggest some promising future research directions.

Motivation & Objective

• To systematically analyze the challenges and solutions in unsupervised person re-identification (Re-ID) for scalable deployment in real-world settings.
• To address the limitations of existing surveys that focus primarily on supervised Re-ID and overlook unsupervised methods.
• To provide in-depth analysis of how recent methods tackle identity label scarcity, pseudo-supervision, cross-camera invariance, and domain shift.
• To summarize and compare evaluation results across leading unsupervised Re-ID methods on standard benchmarks.
• To identify open issues and propose promising future research directions, including end-to-end unsupervised Re-ID and generalizable models.

Proposed method

• The survey conducts a systematic review of deep learning-based unsupervised Re-ID methods, focusing on four key challenges: lack of ground-truth labels, learning discriminative features via pseudo-supervision, achieving cross-camera invariance, and handling domain gaps.
• It analyzes influential methods through the lens of how they address each of the four challenges, evaluating their design choices and trade-offs.
• The authors compare performance across major benchmarks such as Market-1501 and DukeMTMC-Re-ID, highlighting improvements in Rank-1 accuracy over time.
• The survey incorporates insights from recent advances in self-supervised learning, contrastive learning, and domain adaptation to assess solution effectiveness.
• It evaluates both image-based and video-based unsupervised Re-ID methods, emphasizing end-to-end learning and feature alignment techniques.
• The methodology includes a critical synthesis of existing solutions, identifying limitations and opportunities in generalization, scalability, and deployment across domains.

Experimental results

Research questions

• RQ1 How do unsupervised Re-ID methods address the absence of identity labels during training?
• RQ2 What techniques are effective for learning discriminative person features using pseudo-labels in the absence of ground-truth identities?
• RQ3 How do current methods improve cross-camera invariance in person Re-ID without paired supervision?
• RQ4 To what extent can unsupervised models generalize across domains with significant photometric and geometric variations?
• RQ5 What are the key limitations and open challenges in achieving scalable, zero-shot deployment of Re-ID models in real-world settings?

Key findings

• Unsupervised Re-ID performance has significantly improved, with Rank-1 accuracy on Market-1501 rising from 44.4% to 98.5% and on DukeMTMC-Re-ID from 30.8% to over 95% between 2017 and 2020.
• Pseudo-supervision techniques such as clustering and contrastive learning are effective for learning discriminative features without ground-truth labels.
• Cross-camera invariance is improved through data augmentation, self-distillation, and feature alignment, especially when combined with contrastive learning.
• Domain gap mitigation remains a major challenge, with unsupervised domain adaptation methods showing promise but requiring domain-specific adaptation.
• Generalizable Re-ID models that work across diverse, unseen domains are still limited, highlighting the need for pre-training on diverse unlabelled data.
• End-to-end unsupervised Re-ID, particularly leveraging vision transformers, remains underexplored but offers high potential for future scalability.