Source-linked AI summary
Asymmetric Co-Teaching for Unsupervised Cross Domain Person Re-Identification
Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, Shaozi Li
TL;DR
Cross-domain person re-ID must adapt source-trained models to unlabeled target domains despite noisy clustering labels and discarded low-confidence samples. The paper introduces asymmetric co-teaching, in which two models select possibly clean data for each other using diverse and pure sample flows. Experiments show consistent benefits across clustering-based adaptation methods and state-of-the-art adaptation accuracy.
Problem
Source-trained person re-ID models generalize poorly to unseen target domains because of domain shifts.
Method
ACT uses two cooperating models to filter clustering-derived target samples, with one receiving diverse data and the other receiving purer data.
Results
ACT consistently improves adaptation accuracy across clustering-based methods and produces state-of-the-art results on three large-scale datasets.
Takeaways & Limitations
The framework combines possibly clean sample selection with complementary data flows so the adapted models can resist noisy labels while retaining sample diversity.
Abstract
from arXiv · showhide
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.
1 Introduction
Cross-domain person re-ID suffers when source-trained models encounter target-domain shifts. The paper proposes asymmetric co-teaching to filter noisy clustering labels while exposing two models to complementary sample sets, improving adaptation accuracy.
- Source-trained person re-ID models often lose performance on different target domains because of domain shifts.
- Prior unsupervised adaptation methods align domain distributions or discover pseudo labels for unlabeled target images.
- Symmetric co-teaching can repeatedly select easy samples, ignore low-confidence examples, and leave the model vulnerable to local minima.
- ACT separates target samples into inliers and outliers, then uses two models to select small-loss samples for each other through asymmetric data flows.
- Filtered-data training consistently improves adaptation accuracy, while asymmetric co-teaching handles hard samples more effectively than symmetric co-teaching.
- Experiments on three large-scale datasets show that ACT applies across clustering-based methods and achieves state-of-the-art adaptation accuracy.
2 Related Work
Related work covers distribution alignment, clustering-based adaptation, and methods for learning with noisy labels. ACT draws on co-teaching and co-training while addressing limitations of existing approaches for cross-domain re-ID.
- Distribution-alignment methods reduce source-target gaps in image-level or attribute-level spaces, but some require source attribute annotations.
- Clustering-based adaptation creates pseudo labels for target data, but can struggle with choosing cluster counts or handling wrongly labeled samples.
- Noisy-label methods include transition matrices, robust losses, and auxiliary networks that identify or downweight noisy training samples.
- CleanNet requires clean initialization samples, limiting its use in many real-world applications.
- ACT is inspired by co-teaching and co-training, which use two models to help each other resist noisy labels.
3 The Proposed Method
The proposed method adapts a source-trained re-ID model through clustering and asymmetric co-teaching, using two models to select reliable samples from both inliers and outliers. Repeated cooperation aims to combine clean labels with sample diversity, and experiments report strong performance across datasets and adaptation methods.
- Unsupervised domain adaptation combines labeled source data S with unlabeled target data T to learn a model that generalizes on the target testing set.
- ACT uses source-model initialization, clustering-based adaptation, and asymmetric co-teaching as three successive stages.The first stages establish basic discriminability and pseudo labels before the two-model filtering stage.
- ACT trains Mmain and Mco asymmetrically: Mco selects diverse, reliable outlier samples for Mmain, while Mmain selects clean inlier samples for Mco.The models exchange small-loss samples mined from complementary partitions of the target data.
- On Market-1501, DukeMTMC-reID, and CUHK03, the proposed algorithm outperforms image-level, attribute-level, clustering-based, and hybrid methods by a large margin.
- DBSCAN divides target samples into inliers Ti and outliers To, assigns pseudo labels to outliers through nearest inliers, and avoids discarding them.The baseline adaptation fine-tunes on clustered inliers, whereas ACT explicitly reuses outliers.
- The method repeatedly alternates inlier/outlier generation and asymmetric co-teaching, then uses the trained Mmain as the final adapted model.For co-teaching, the method selects the smallest K% triplet-loss anchors rather than applying the original symmetric strategy directly.
4 Experiment
Experiments evaluate ACT across benchmarks, clustering methods, and controlled variants. Results show improved adaptation accuracy, useful small-loss sample selection, and progressively better clustering quality.
- Experimental Setup: ACT is evaluated on Market-1501, DukeMTMC-reID, and CUHK03 using mAP and rank-1 accuracy.The adaptation uses clustering-based training and asymmetric co-teaching stages.
- Comparison with State-of-the-Arts: ACT outperforms competing methods across three large-scale benchmarks and achieves approximately 32% and 27% mAP improvements over MMFA and CFSM on D→M and M→D.The reported comparisons cover multiple adaptation tasks.
- Sample Mining: Small-loss samples selected from inliers provide relatively reliable, diverse training images, while many outliers are high-variance samples affected by occlusion and illumination.These selected samples are used to train the main network during asymmetric co-teaching.
- Ablation Study: ACT achieves the highest mAP scores in the ablation study, reaching 60.6% and 54.5% on the two adaptation tasks.The ablation also finds that incorporating outliers improves mAP by 2.7% and 0.7%, while asymmetric co-teaching performs best overall.
- Training Dynamics: During adaptation, F-scores increase across iterations while the number of outliers decreases, indicating improving clustering quality.The collaborator model remains weaker because it is trained mainly on easy small-loss samples.
- Variant Evaluation: ACT remains effective with k-means, achieving 3.3% and 3.1% improvements over vanilla k-means-based adaptation.The experiments treat the furthest samples from centroids as outliers because k-means does not directly generate them.
5 Conclusion
The proposed asymmetric co-training framework adapts person re-ID across domains by having two networks select possibly clean target samples for each other while receiving samples with different purity and diversity.
- Two networks, initialized with the same weights, are named the Main Model and Collaborator Model.
- The models select possibly clean target samples for each other to help the adapted Main Model resist noisy labels.
- Asymmetric data flows make Main Model samples as diverse as possible and Collaborator Model samples as pure as possible.
- The proposed method works on three large-scale datasets and is proposed for future application to other unsupervised domain adaptation tasks.