Uncertainty-aware Clustering for Unsupervised Domain Adaptive Object Re-identification

Pengfei Wang, Changxing Ding, Wentao Tan, Mingming Gong, Kui Jia, Dacheng Tao

Abstract

Unsupervised Domain Adaptive (UDA) object re-identification (Re-ID) aims at adapting a model trained on a labeled source domain to an unlabeled target domain. State-of-the-art object Re-ID approaches adopt clustering algorithms to generate pseudo-labels for the unlabeled target domain. However, the inevitable label noise caused by the clustering procedure significantly degrades the discriminative power of Re-ID model. To address this problem, we propose an uncertainty-aware clustering framework (UCF) for UDA tasks. First, a novel hierarchical clustering scheme is proposed to promote clustering quality. Second, an uncertainty-aware collaborative instance selection method is introduced to select images with reliable labels for model training. Combining both techniques effectively reduces the impact of noisy labels. In addition, we introduce a strong baseline that features a compact contrastive loss. Our UCF method consistently achieves state-of-the-art performance in multiple UDA tasks for object Re-ID, and significantly reduces the gap between unsupervised and supervised Re-ID performance. In particular, the performance of our unsupervised UCF method in the MSMT17 → Market1501 task is better than that of the fully supervised setting on Market1501. The code of UCF is available at https://github.com/Wang-pengfei/UCF.

Introduction

Object Re-identification (Re-ID) is the task of identifying the same person or vehicle across different non-overlapping camera views. While supervised models perform well, they often fail when deployed in new environments (the "target domain") due to domain gaps—differences in lighting, camera angles, and backgrounds.

To address this without manual labels, Unsupervised Domain Adaptation (UDA) is employed. A common strategy involves clustering target-domain features to generate "pseudo-labels" for training. However, this creates a "chicken-and-egg" problem: an imperfect model produces noisy clusters, and training on these noisy labels prevents the model from reaching peak performance. Most existing methods either ignore this noise or discard "unreliable" clusters entirely, wasting valuable data.

The Uncertainty-aware Clustering Framework (UCF)

The proposed Uncertainty-aware Clustering Framework (UCF) tackles the noisy label problem through a combination of hierarchical clustering and collaborative instance selection.

Overall architecture of the UCF framework, showing the dual-network pipeline and the integration of HC and UCIS modules.

Hierarchical Clustering (HC)

Instead of treating "loose" clusters as outliers and discarding them, UCF uses the silhouette coefficient to measure cluster tightness. If a cluster is deemed unreliable, UCF performs a second, finer-grained clustering pass within that specific group. This "decomposition over discarding" approach recognizes that unreliable clusters often contain multiple valid identities that the model isn't yet strong enough to separate at a coarse level.

Uncertainty-aware Collaborative Instance Selection (UCIS)

UCF maintains two versions of the model: the current "Net" and a "mean-Net" (a temporally averaged version). By comparing how both models cluster a specific image, UCF calculates an uncertainty score. If an image ends up in highly similar clusters across both models, its pseudo-label is considered reliable and preserved for training. This uses temporal consistency as a proxy for label reliability.

Detailed mechanisms of (a) Hierarchical Clustering, (b) Uncertainty-aware Collaborative Instance Selection, and (c) the overall noisy label filtering process.

Compact Contrastive Loss

Standard contrastive losses require storing feature vectors for every image, which is memory-intensive. UCF introduces a class-level memory bank that stores only one feature vector per cluster. This efficiency allows for more frequent updates and improved accuracy.

Experimental Evaluation

The authors evaluated UCF on a comprehensive suite of benchmarks, including real-world person Re-ID (Market-1501, DukeMTMC-ReID, MSMT17), synthetic person Re-ID (PersonX), and vehicle Re-ID (VeRi-776, VehicleID, VehicleX).

Key Results

  • State-of-the-Art Performance: UCF consistently outperformed previous leaders like MMT and SpCL. On the MSMT17 → Market1501 task, UCF achieved 85.5% mAP, a significant jump over MMT's 75.6%.
  • Surpassing the "Oracle": Remarkably, in the MSMT17 → Market1501 task, the unsupervised UCF method outperformed the fully supervised "Oracle" setting on Market1501. This suggests that pre-training on a larger, more diverse dataset and adapting carefully can be more powerful than training on a smaller labeled dataset from scratch.
  • Robustness to Domain Shifts: On the PersonX → MSMT17 task (synthetic-to-real), UCF improved mAP from 22.7% to 28.3%, demonstrating superior handling of extreme domain gaps.
  • Efficiency Gains: The class-level memory bank reduced memory costs while improving the accuracy of the contrastive loss by 2.5% mAP compared to instance-level approaches.

Conclusion

UCF narrows the gap between unsupervised and supervised Re-ID by intelligently handling label noise rather than simply avoiding it. While a performance delta still exists on very large-scale datasets like MSMT17, the framework's ability to decompose unreliable clusters and leverage temporal consistency marks a significant step forward. Future work may explore parameter-free clustering methods to further reduce sensitivity to initial hyperparameters like the DBSCAN neighbor distance.

Summary updated

Abstract

Unsupervised Domain Adaptive (UDA) object re-identification (Re-ID) aims at adapting a model trained on a labeled source domain to an unlabeled target domain. State-of-the-art object Re-ID approaches adopt clustering algorithms to generate pseudo-labels for the unlabeled target domain. However, the inevitable label noise caused by the clustering procedure significantly degrades the discriminative power of Re-ID model. To address this problem, we propose an uncertainty-aware clustering framework (UCF) for UDA tasks. First, a novel hierarchical clustering scheme is proposed to promote clustering quality. Second, an uncertainty-aware collaborative instance selection method is introduced to select images with reliable labels for model training. Combining both techniques effectively reduces the impact of noisy labels. In addition, we introduce a strong baseline that features a compact contrastive loss. Our UCF method consistently achieves state-of-the-art performance in multiple UDA tasks for object Re-ID, and significantly reduces the gap between unsupervised and supervised Re-ID performance. In particular, the performance of our unsupervised UCF method in the MSMT17$\to$Market1501 task is better than that of the fully supervised setting on Market1501. The code of UCF is available at https://github.com/Wang-pengfei/UCF.
BibTeX
@article{wang2023uncertainty,
  author = {Pengfei Wang and Changxing Ding and Wentao Tan and Mingming Gong and Kui Jia and Dacheng Tao},
  title = {{Uncertainty-aware Clustering for Unsupervised Domain Adaptive Object Re-identification}},
  journal = {IEEE Trans. Multimedia},
  year = {2023},
  url = {https://openreview.net/forum?id=IiBX9MEnzC},
}
Figure