Unsupervised Domain Adaptation in Person re-ID via k-Reciprocal Clustering and Large-Scale Heterogeneous Environment Synthesis

Devinder Kumar, Parthipan Siva, Paul Marchwica, Alexander Wong

Abstract An ongoing major challenge in computer vision is the task of person re-identification, where the goal is to match individuals across different, non-overlapping camera views. While recent success has been achieved via supervised learning using deep neural networks, such methods have limited widespread adoption due to the need for large-scale, customized data annotation. As such, there has been a recent focus on unsupervised learning approaches to mitigate the data annotation issue; however, current approaches in literature have limited performance compared to supervised learning approaches as well as limited applicability for adoption in new environments. In this paper, we address the aforementioned challenges faced in person re-identification for real-world, practical scenarios by introducing a novel, unsupervised domain adaptation approach for person re-identification. This is accomplished through the introduction of: i) k-reciprocal tracklet Clustering for Unsupervised Domain Adaptation (ktCUDA) (for pseudo-label generation on target domain), and ii) Synthesized Heterogeneous RE-id Domain (SHRED) composed of large-scale heterogeneous independent source environments (for improving robustness and adaptability to a wide diversity of target environments). Experimental results across four different image and video benchmark datasets show that the proposed ktCUDA and SHRED approach achieves an average improvement of +5.7 mAP in re-identification performance when compared to existing state-of-the-art methods, as well as demonstrate better adaptability to different types of environments.

Introduction

Matching individuals across a network of non-overlapping cameras—known as Person re-identification (re-ID)—is a cornerstone of modern surveillance and retail analytics. However, the "supervised learning" gold standard requires thousands of manually labeled images for every new deployment. Most re-ID systems suffer from a "domain gap" where a model trained on one set of cameras (the source) fails in a new environment (the target) due to changes in lighting, camera angles, and backgrounds.

Current Unsupervised Domain Adaptation (UDA) solutions face two major bottlenecks: the Metric Gap, where simple Euclidean distance creates noisy labels, and the Source Gap, where models pre-trained on narrow datasets lack robustness. This paper proposes a two-pronged solution: a smarter clustering algorithm called ktCUDA and a massive, diverse "super-dataset" called SHRED.

Overview of the proposed UDA framework, illustrating the transition from source domain pre-training to target domain adaptation via ktCUDA clustering.

Methodology

1. ktCUDA (k-Reciprocal Tracklet Clustering)

Instead of relying on simple proximity in feature space, ktCUDA utilizes k-reciprocal neighbors. The core logic is that if Image A is the top neighbor of Image B, and Image B is the top neighbor of Image A, there is a high probability they represent the same person. The model builds a directed graph of these mutual relationships and thresholds it to form clusters. These clusters act as "pseudo-labels," allowing the model to fine-tune itself on target data without human intervention.

2. SHRED (Synthesized Heterogeneous RE-id Domain)

To provide a robust starting point, the authors created SHRED by merging 10 different datasets (including MSMT17 and CUHK03) while filtering out overlapping identities. This heterogeneous "super-source" contains a vast variety of lighting, poses, and backgrounds, ensuring the initial model is highly generalizable.

Experimental Results

The combination of SHRED and ktCUDA consistently outperformed state-of-the-art (SOTA) methods across five major benchmarks:

  • Market-1501: Achieved 49.4% mAP, an 8.2% improvement over previous SOTA.
  • CUHK03: Achieved 41.6% mAP, a 10.4% improvement.
  • Average Gain: Across all tested datasets, the method showed an average improvement of +5.7 mAP.

Key insights from the results suggest that reciprocity is a stronger signal than proximity for filtering false positives during clustering, and that environmental diversity in the source domain is more critical than sheer data volume.

Limitations and Future Work

Despite its robustness, ktCUDA can struggle with clustering errors, such as "Mixed Clusters" (grouping different people) or splitting a single individual into multiple clusters.

Example of a clustering challenge where a single individual is split into two distinct clusters (Cluster 29 and Cluster 27), which can confuse the training process.

Future work aims to develop "noise-aware" loss functions that can identify and ignore these imperfect clusters during fine-tuning, further improving the performance of fully autonomous re-ID systems.

Summary updated

Abstract

An ongoing major challenge in computer vision is the task of person re-identification, where the goal is to match individuals across different, non-overlapping camera views. While recent success has been achieved via supervised learning using deep neural networks, such methods have limited widespread adoption due to the need for large-scale, customized data annotation. As such, there has been a recent focus on unsupervised learning approaches to mitigate the data annotation issue; however, current approaches in literature have limited performance compared to supervised learning approaches as well as limited applicability for adoption in new environments. In this paper, we address the aforementioned challenges faced in person re-identification for real-world, practical scenarios by introducing a novel, unsupervised domain adaptation approach for person re-identification. This is accomplished through the introduction of: i) k-reciprocal tracklet Clustering for Unsupervised Domain Adaptation (ktCUDA) (for pseudo-label generation on target domain), and ii) Synthesized Heterogeneous RE-id Domain (SHRED) composed of large-scale heterogeneous independent source environments (for improving robustness and adaptability to a wide diversity of target environments). Experimental results across four different image and video benchmark datasets show that the proposed ktCUDA and SHRED approach achieves an average improvement of +5.7 mAP in re-identification performance when compared to existing state-of-the-art methods, as well as demonstrate better adaptability to different types of environments.
BibTeX
@inproceedings{kumar2020unsupervised,
  author = {Devinder Kumar and Parthipan Siva and Paul Marchwica and Alexander Wong},
  title = {{Unsupervised Domain Adaptation in Person re-ID via k-Reciprocal Clustering and Large-Scale Heterogeneous Environment Synthesis}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
  url = {https://openreview.net/forum?id=5hV6lSlxmL},
}
Figure