[Paper Review] Asymmetric Co-Teaching for Unsupervised Cross Domain Person Re-Identification

Fengxiang Yang, Ke Li+8 more|arXiv (Cornell University)|Dec 3, 2019

Video Surveillance and Tracking Methods31 references13 citations

TL;DR

This paper proposes Asymmetric Co-Teaching (ACT), a novel framework for unsupervised cross-domain person re-identification that improves clustering-based domain adaptation by using two asymmetric models: one trained on diverse inliers and outliers, and another on pure inliers to resist noisy pseudo-labels. ACT achieves state-of-the-art mAP of 60.6% and 54.5% on Duke-to-Market and Market-to-Duke benchmarks, respectively, by iteratively refining clean and diverse training samples.

ABSTRACT

Person re-identification (re-ID), is a challenging task due to the high variance within identity samples and imaging conditions. Although recent advances in deep learning have achieved remarkable accuracy in settled scenes, i.e., source domain, few works can generalize well on the unseen target domain. One popular solution is assigning unlabeled target images with pseudo labels by clustering, and then retraining the model. However, clustering methods tend to introduce noisy labels and discard low confidence samples as outliers, which may hinder the retraining process and thus limit the generalization ability. In this study, we argue that by explicitly adding a sample filtering procedure after the clustering, the mined examples can be much more efficiently used. To this end, we design an asymmetric co-teaching framework, which resists noisy labels by cooperating two models to select data with possibly clean labels for each other. Meanwhile, one of the models receives samples as pure as possible, while the other takes in samples as diverse as possible. This procedure encourages that the selected training samples can be both clean and miscellaneous, and that the two models can promote each other iteratively. Extensive experiments show that the proposed framework can consistently benefit most clustering-based methods, and boost the state-of-the-art adaptation accuracy. Our code is available at https://github.com/FlyingRoastDuck/ACT_AAAI20.

Motivation & Objective

• To address noisy pseudo-labels and discarded low-confidence samples in clustering-based unsupervised domain adaptation for person re-identification.
• To overcome the limitations of symmetric co-teaching, which ignores hard, low-confidence samples due to their high training loss.
• To improve model generalization by explicitly filtering and reusing outliers and inliers through asymmetric data flow.
• To enhance the discriminative ability of both models through iterative mutual promotion using clean and diverse training data.

Proposed method

• The framework initializes two models, Main and Collaborator, from a pre-trained source model fine-tuned on target pseudo-labels.
• It divides target samples into inliers (high-confidence clusters) and outliers (low-confidence, high-variation samples) using clustering (e.g., DBSCAN or k-means with outlier thresholding).
• The Main model trains on all inliers and outliers to maximize diversity, while the Collaborator model trains only on inliers with small loss values to ensure purity.
• During co-teaching, each model selects samples with minimal loss from the other’s output, enabling mutual refinement and noise resistance.
• The process iteratively improves clustering quality and model performance by retraining with increasingly clean and representative samples.
• The method is compatible with multiple clustering methods, including k-means and DBSCAN, by defining outliers as farthest samples from cluster centroids.

Experimental results

Research questions

• RQ1 Can asymmetric co-teaching improve the robustness of clustering-based unsupervised domain adaptation in person re-identification?
• RQ2 How does the inclusion of low-confidence (outlier) samples affect model generalization when combined with noise-resistant training?
• RQ3 Does the asymmetric data flow—diverse inputs for one model and pure inputs for the other—lead to better performance than symmetric co-teaching?
• RQ4 To what extent is the proposed framework compatible with different clustering algorithms such as k-means and DBSCAN?
• RQ5 Does the iterative refinement of pseudo-labels and model weights lead to consistent improvement in mAP and rank-1 accuracy?

Key findings

• ACT achieves a new state-of-the-art mAP of 60.6% on Duke-to-Market and 54.5% on Market-to-Duke, outperforming baseline clustering methods.
• The method improves mAP by 3.3% and 3.1% over vanilla k-means-based adaptation when using 20% and 30% of the farthest samples as outliers.
• F-score for clustering quality increases over iterations, indicating improved cluster discriminativeness and model generalization.
• The number of outliers decreases over training, confirming that the model learns to better assign and filter hard samples.
• The Collaborator model consistently achieves lower performance than the Main model, confirming its role in learning from pure, low-loss samples.
• ACT maintains consistent gains across different clustering methods, demonstrating strong compatibility and generalization.