Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification (2001.01526v2)

Abstract: Person re-identification (re-ID) aims at identifying the same persons' images across different cameras. However, domain diversities between different datasets pose an evident challenge for adapting the re-ID model trained on one dataset to another one. State-of-the-art unsupervised domain adaptation methods for person re-ID transferred the learned knowledge from the source domain by optimizing with pseudo labels created by clustering algorithms on the target domain. Although they achieved state-of-the-art performances, the inevitable label noise caused by the clustering procedure was ignored. Such noisy pseudo labels substantially hinders the model's capability on further improving feature representations on the target domain. In order to mitigate the effects of noisy pseudo labels, we propose to softly refine the pseudo labels in the target domain by proposing an unsupervised framework, Mutual Mean-Teaching (MMT), to learn better features from the target domain via off-line refined hard pseudo labels and on-line refined soft pseudo labels in an alternative training manner. In addition, the common practice is to adopt both the classification loss and the triplet loss jointly for achieving optimal performances in person re-ID models. However, conventional triplet loss cannot work with softly refined labels. To solve this problem, a novel soft softmax-triplet loss is proposed to support learning with soft pseudo triplet labels for achieving the optimal domain adaptation performance. The proposed MMT framework achieves considerable improvements of 14.4%, 18.2%, 13.1% and 16.4% mAP on Market-to-Duke, Duke-to-Market, Market-to-MSMT and Duke-to-MSMT unsupervised domain adaptation tasks. Code is available at https://github.com/yxgeee/MMT.

Overview of Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification

This paper presents an innovative approach for addressing the challenges in unsupervised domain adaptation (UDA) in person re-identification (re-ID). Specifically, it introduces the Mutual Mean-Teaching (MMT) framework designed to mitigate the adverse effects of noisy pseudo labels generated by clustering algorithms, which are common in state-of-the-art UDA methods for re-ID.

Methodology

The paper identifies the problem of inevitable label noise within pseudo labels as a substantial hindrance to improving feature representations on the target domain. Through an unsupervised framework, MMT innovatively refines pseudo labels in an alternative training manner, utilizing both off-line refined hard labels and on-line refined soft labels.

Key Components:

1. Collaborative Training Mechanism: MMT employs a teacher-student philosophy, training two networks simultaneously. The networks produce on-line refined pseudo labels, leveraging past temporal averages to avoid bias amplification. This temporal ensembling encourages reliable label generation, enhancing feature learning.
2. Soft Softmax-Triplet Loss: The framework proposes a novel soft softmax-triplet loss compatible with soft pseudo labels, critical for effective domain adaptation. This loss addresses the triplet loss's limitations, which cannot naturally accommodate soft labels, thereby improving UDA performance.
3. Pseudo Label Refinery: MMT conducts pseudo label refinement by balancing hard and soft pseudo labels, ensuring robust learning despite the intrinsic noise of clustering-generated labels.

Results

The proposed MMT framework demonstrates considerable improvements across multiple person re-ID datasets:

• Achieves notable mAP increases of 14.4%, 18.2%, 13.4%, and 16.4% on Market-to-Duke, Duke-to-Market, Market-to-MSMT, and Duke-to-MSMT tasks, respectively.
• The performance enhancements underscore the effectiveness of mitigating noisy pseudo labels and refining learning processes via MMT.

Implications

The findings have significant implications for the field of computer vision, particularly in enhancing the robustness and accuracy of re-ID systems under unsupervised conditions. The methodology adapts effectively across different camera domains, providing a versatile solution applicable to various scenarios in automated surveillance and security systems.

Future Directions

The research opens avenues for exploring deeper integration of temporal ensemble models in collaborative learning contexts, potentially extending beyond re-ID applications. Additionally, further refinement of soft triplet loss functions could provide even more substantial gains in various domain adaptation challenges.

In conclusion, the Mutual Mean-Teaching framework provides meaningful advancements in unsupervised domain adaptation for person re-ID, addressing core challenges of noisy pseudo labels and proposing effective solutions that yield significant performance improvements.