Unsupervised Domain Adaptation with Noise Resistible Mutual-Training for Person Re-identification

Fang Zhao, Shengcai Liao, Guo-Sen Xie, Jian Zhao, Kaihao Zhang, Ling Shao

Unsupervised Domain Adaptation with Noise Resistible Mutual-Training for Person Re-identification

Person re-identification (re-ID) aims to match individuals across non-overlapping camera views. While supervised methods have achieved high accuracy, they often suffer from domain shift when applied to new environments—differences in lighting, camera angles, and backgrounds can cause performance to plummet. Unsupervised Domain Adaptation (UDA) attempts to bridge this gap by training on unlabeled target data, typically through "self-training" where the model clusters data to generate pseudo-labels. However, clustering is imperfect, and these noisy pseudo-labels can lead to error accumulation, where the model memorizes incorrect patterns.

The NRMT Framework

To address the issue of label noise, this research introduces Noise Resistible Mutual-Training (NRMT). The core idea is to maintain two deep networks, $f$ and $g$, that provide mutual supervision, effectively "peer-reviewing" each other's learning process.

The NRMT framework architecture featuring collaborative clustering and mutual instance selection.

The framework consists of two primary components:

  1. Collaborative Clustering: Instead of a single network relying on its own potentially erroneous labels, NRMT uses two networks that perform clustering independently. During training, each network optimizes a joint loss function that incorporates pseudo-labels from both itself and its peer. This approach exploits the "memorization effect," where networks learn clean patterns before overfitting to noise.
  2. Mutual Instance Selection: This mechanism filters the training data within each mini-batch. The networks select image triplets based on Peer-Confidence (ensuring the relationship is reliable according to the peer network) and Relationship Disagreement (prioritizing samples that are informative and challenging).
Comparison of noisy, easy, and informative triplets in the feature spaces of networks f and g.

As illustrated in the figure above, NRMT distinguishes between different types of data relationships:

  • (a) Noisy triplets: Where networks disagree significantly on the identity or relationship, suggesting potential errors.
  • (b) Reliable but easy triplets: Which are correctly identified but provide little new information for the model to learn from.
  • (c) Reliable and informative triplets: These are "hard" samples that are likely correct but challenging, driving the learning process forward.

Experimental Evaluation

NRMT was evaluated on major re-ID benchmarks, including Market-1501, DukeMTMC-reID, and the large-scale MSMT17. In cross-domain scenarios, NRMT demonstrated substantial improvements over existing state-of-the-art (SOTA) methods:

  • Duke $\to$ Market-1501: Achieved 71.7% mAP, significantly outperforming previous benchmarks like ACT (60.6%) and SSG (58.3%).
  • Market $\to$ Duke: Reached 62.2% mAP, an 8% lead over the next best method.
  • MSMT17 Transfer: On this challenging dataset, NRMT nearly doubled the mAP of previous benchmarks (20.6% vs. 13.3% for SSG).

Discussion and Future Directions

NRMT offers a powerful solution to the label noise problem in unsupervised person re-ID. By utilizing mutual training and a sophisticated instance selection strategy, it identifies reliable and informative data points, preventing the model from stagnating on easy samples or crashing due to noise.

While the method introduces additional computational costs by maintaining two networks, its performance gains make it a significant advancement. Future work may explore combining this mutual-training framework with other UDA techniques, such as camera-aware adaptation, to further push the boundaries of unsupervised re-identification.

Summary updated

Abstract

Unsupervised domain adaptation (UDA) in the task of person re-identification (re-ID) is highly challenging due to large domain divergence and no class overlap between domains. Pseudo-label based self-training is one of the representative techniques to address UDA. However, label noise caused by unsupervised clustering is always a trouble to self-training methods. To depress noises in pseudo-labels, this paper proposes a Noise Resistible Mutual-Training (NRMT) method, which maintains two networks during training to perform collaborative clustering and mutual instance selection. On one hand, collaborative clustering eases the fitting to noisy instances by allowing the two networks to use pseudo-labels provided by each other as an additional supervision. On the other hand, mutual instance selection further selects reliable and informative instances for training according to the peer-confidence and relationship disagreement of the networks. Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art UDA methods for person re-ID.
BibTeX
@inproceedings{zhao2020unsupervised,
  author = {Fang Zhao and Shengcai Liao and Guo-Sen Xie and Jian Zhao and Kaihao Zhang and Ling Shao},
  title = {{Unsupervised Domain Adaptation with Noise Resistible Mutual-Training for Person Re-identification}},
  booktitle = {European Conference on Computer Vision},
  year = {2020},
  doi = {10.1007/978-3-030-58621-8\_31},
}
Figure