Online Pseudo Label Generation by Hierarchical Cluster Dynamics for Adaptive Person Re-identification

Yi Zheng, Shixiang Tang, Guolong Teng, Yixiao Ge, Kaijian Liu, Jing Qin, Donglian Qi, Dapeng ChenICCV 2021 · OpenReview

Abstract

Adaptive person re-identification (adaptive ReID) targets at transferring learned knowledge from the labeled source domain to the unlabeled target domain. Pseudo-label-based methods that alternatively generate pseudo labels and optimize the training model have demonstrated great effectiveness in this field. However, the generated pseudo labels are inaccurate and cannot reflect the true semantic meaning of the unlabeled samples. We consider such inaccuracy stems from both the lagged update of the pseudo labels as well as the simple criterion of the employed clustering method. To tackle the problem, we propose an online pseudo label generation by hierarchical cluster dynamics for adaptive ReID. In particular, hierarchical label banks are constructed for all the samples in the dataset, and we update the pseudo labels of the sample in each coming mini-batch, performing the model optimization and the label generation simultaneously. A new hierarchical cluster dynamics is built for the label update, where cluster merge and cluster split are driven by a possibility computed by the label propagation. Our method can achieve better pseudo labels and higher reid accuracy. Extensive experiments on Market-to-Duke, Duke-to-Market, MSMT-to-Market, MSMT-to-Duke, Market-to-MSMT, and Duke-to-MSMT verify the effectiveness of our proposed method.

Introduction

Person Re-identification (ReID) is the task of identifying the same individual across different non-overlapping camera views. While deep learning has made massive strides here, models often fail when moved to a new environment (the "target domain") because of domain shift—differences in lighting, camera angles, and backgrounds. Adaptive ReID tries to fix this by transferring knowledge from a labeled source dataset to an unlabeled target dataset. The standard approach is to use "pseudo labels": the model clusters unlabeled images in the target domain and treats those clusters as ground-truth identities.

However, existing methods have two major flaws: the "lag" caused by offline clustering once per epoch, and rigid clustering criteria that ignore varying identity difficulty. This paper introduces an online framework that generates pseudo labels on-the-fly using hierarchical cluster dynamics.

Proposed Method

Overview of the online pseudo label generation framework.

The core of the approach rests on two pillars: a hierarchical label bank and dynamic cluster adjustment.

Hierarchical Label Bank

Instead of a single set of labels, the system maintains a hierarchical label bank with $H$ levels. This structure follows a "cluster preserve property," meaning if two samples are grouped together at a fine-grained level, they must stay together at coarser levels. This bottom-up approach helps the model capture complex feature distributions, from simple viewpoint similarities to cross-posture identity matching.

Hierarchical Cluster Dynamics

In each level of the hierarchy, the model performs two operations: Cluster Split and Cluster Merge. Unlike traditional hierarchical methods that only allow merging, this bi-directional adjustment allows the model to correct early mistakes.

Conceptual illustration of hierarchical cluster dynamics across levels.

These decisions are driven by Label Propagation. The model calculates the probability of two units belonging to the same class based on their k-nearest neighbors in the feature space. If the probability exceeds a threshold, they merge; if it drops, they split. This online update per mini-batch allows pseudo labels to evolve alongside the feature extractor, preventing overfitting to noisy, outdated labels.

Experimental Results

The researchers evaluated their method on industry-standard benchmarks: Market-1501, DukeMTMC-reID, and MSMT17. The proposed method consistently outperformed state-of-the-art (SOTA) methods like SpCL and MMT:

  • Adaptive ReID: On the Duke-to-Market task, it achieved 80.0% mAP (mean Average Precision), outperforming the previous best by 1.9%.
  • Pure Unsupervised ReID: Without source labels, the method outperformed the BUC method by a staggering 39.8% mAP on Market-1501.
  • Robustness: Gains were most pronounced on the difficult MSMT17 dataset, demonstrating the ability to handle large-scale, complex data.

Conclusion

By introducing online updates and a split-merge dynamic within a hierarchy, this method significantly improves the accuracy of pseudo labels in adaptive ReID. While performance gains plateaued beyond two levels in current datasets, the framework provides a robust foundation for handling more diverse, "in-the-wild" data in future research.

Summary updated

Abstract

Adaptive person re-identification (adaptive ReID) targets at transferring learned knowledge from the labeled source domain to the unlabeled target domain. Pseudo-label-based methods that alternatively generate pseudo labels and optimize the training model have demonstrated great effectiveness in this field. However, the generated pseudo labels are inaccurate and cannot reflect the true semantic meaning of the unlabeled samples. We consider such inaccuracy stems from both the lagged update of the pseudo labels as well as the simple criterion of the employed clustering method. To tackle the problem, we propose an online pseudo label generation by hierarchical cluster dynamics for adaptive ReID. In particular, hierarchical label banks are constructed for all the samples in the dataset, and we update the pseudo labels of the sample in each coming mini-batch, performing the model optimization and the label generation simultaneously. A new hierarchical cluster dynamics is built for the label update, where cluster merge and cluster split are driven by a possibility computed by the label propagation. Our method can achieve better pseudo labels and higher reid accuracy. Extensive experiments on Market-to-Duke, Duke-to-Market, MSMT-to-Market, MSMT-to-Duke, Market-to-MSMT, and Duke-to-MSMT verify the effectiveness of our proposed method.
BibTeX
@inproceedings{zheng2021online,
  author = {Yi Zheng and Shixiang Tang and Guolong Teng and Yixiao Ge and Kaijian Liu and Jing Qin and Donglian Qi and Dapeng Chen},
  title = {{Online Pseudo Label Generation by Hierarchical Cluster Dynamics for Adaptive Person Re-identification}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  url = {https://openreview.net/forum?id=7EV5KAVFDEM},
}
Figure