Source-linked AI summary
Exploiting Sample Uncertainty for Domain Adaptive Person Re-Identification
Kecheng Zheng, Cuiling Lan, Wenjun Zeng, Zhizheng Zhang, Zheng-Jun Zha
TL;DR
UDA person ReID is challenged by noisy clustering pseudo-labels caused by domain gaps, which can mislead feature learning. This paper introduces UNRN, using mean-teacher consistency to estimate pseudo-label reliability and re-weight ID, triplet, and contrastive losses. The method achieves state-of-the-art performance across benchmark datasets, while ReID deployment raises privacy concerns requiring regulation.
Problem
Domain gaps make clustering-based target pseudo-labels noisy, which can mislead feature representation learning and deteriorate UDA person ReID performance.
Method
UNRN uses a temporally averaged mean-teacher model to estimate pseudo-label reliability from student–teacher consistency and re-weight ID, triplet, and contrastive losses.
Results
UNRN achieves state-of-the-art performance on all benchmark datasets and significantly outperforms the second-best UDA methods across four domain-adaptation settings.
Takeaways & Limitations
Uncertainty-guided optimization softly suppresses error-prone pseudo-label samples while exploiting their contributions across multiple ReID losses.
Takeaways & Limitations
Deploying person ReID can violate human privacy, requiring strict regulations and laws to protect data and ensure legal use.
Abstract
from arXiv · showhide
Many unsupervised domain adaptive (UDA) person re-identification (ReID) approaches combine clustering-based pseudo-label prediction with feature fine-tuning. However, because of domain gap, the pseudo-labels are not always reliable and there are noisy/incorrect labels. This would mislead the feature representation learning and deteriorate the performance. In this paper, we propose to estimate and exploit the credibility of the assigned pseudo-label of each sample to alleviate the influence of noisy labels, by suppressing the contribution of noisy samples. We build our baseline framework using the mean teacher method together with an additional contrastive loss. We have observed that a sample with a wrong pseudo-label through clustering in general has a weaker consistency between the output of the mean teacher model and the student model. Based on this finding, we propose to exploit the uncertainty (measured by consistency levels) to evaluate the reliability of the pseudo-label of a sample and incorporate the uncertainty to re-weight its contribution within various ReID losses, including the identity (ID) classification loss per sample, the triplet loss, and the contrastive loss. Our uncertainty-guided optimization brings significant improvement and achieves the state-of-the-art performance on benchmark datasets.
1 Introduction
UDA person ReID uses labeled source and unlabeled target data but struggles with domain gaps and noisy clustering pseudo-labels. UNRN estimates pseudo-label reliability from student–teacher consistency and re-weights noisy samples, achieving state-of-the-art benchmark performance.
- Motivation: UDA person ReID exploits labeled source and unlabeled target data to address poor cross-domain generalization.The task matches person images across times, spaces, and cameras, with applications including smart retail tracking and image retrieval.
- Motivation: Typical UDA pipelines pre-train on source labels, cluster unlabeled target data for pseudo-labels, and fine-tune features iteratively.Clustering pseudo-labels can be noisy because of source–target divergence and imperfect clustering, misleading feature learning and harming adaptation.
- Approach: The paper measures uncertainty through inconsistency between student and mean-teacher output features, observing higher uncertainty for wrong or noisy pseudo-labels.The observation uses target samples from Duke→Market and compares red noisy-label and green clean-label curves.
- Approach: UNRN estimates pseudo-label credibility and re-weights sample contributions in ID classification, triplet, and contrastive losses.Its uncertainty estimation calculates inconsistency between two models’ predicted soft multilabels.
- Results: UNRN achieves state-of-the-art performance on unsupervised person ReID benchmark datasets.The introduction reports extensive experiments on benchmark datasets and effectiveness of both the framework and its components.
2 Related Work
Related UDA ReID methods use clustering, domain translation, memory banks, or uncertainty modeling to exploit unlabeled target data. UNRN differs by softly weighting uncertainty-based pseudo-label reliability with a mean-teacher framework rather than relying on multiple mutually learning networks or hard instance selection.
- Clustering-based methods: Clustering-based UDA ReID methods generate hard or soft pseudo-labels from target-data clusters and fine-tune models with those labels.These methods are widely used because of their superior performance.
- Noise suppression: Mutual-learning methods refine soft pseudo-labels, while NRMT reduces noisy-instance fitting through collaborative clustering and mutual instance selection.NRMT’s hard reliable-instance selection requires threshold tuning and may discard useful information.
- UNRN: UNRN estimates pseudo-label reliability through uncertainty and softly exploits it in ReID losses without requiring two networks for mutual learning.Its mean-teacher model is a temporally averaged base network used to estimate each target sample’s uncertainty.
- Domain translation: Domain translation methods transfer labeled source images into target style, but unsatisfactory translated-image quality hinders their advancement.The transferred images retain inherited ground-truth labels for fine-tuning.
- Memory-bank methods: Memory-bank methods support contrastive learning by exploiting global similarities, mining hard negatives across batches, or enforcing invariances.The cited approaches use memory banks for instance similarity, hard-negative mining, and exemplar, camera, or neighborhood invariance.
- Uncertainty modeling: Uncertainty modeling has been explored for data-dependent uncertainty, model uncertainty, and uncertainty on annotation to suppress training noise.These directions include uncertainty estimation in both supervised and other learning settings.
3 Uncertainty-guided Noise Resilient Network
UNRN addresses noisy target-domain pseudo-labels by estimating sample uncertainty from teacher–student output inconsistency and re-weighting ReID losses. It builds on a clustering-based mean-teacher baseline and uses uncertainty-guided optimization across classification, triplet, and contrastive objectives.
- Strong baseline: UNRN combines source-data fine-tuning, contrastive loss, and a mean teacher that supplies features for target clustering and final inference.The teacher is a temporally averaged student model with no gradient back-propagation.
- Framework: UDA ReID adapts through source pre-training, target clustering, and pseudo-label fine-tuning, but domain gaps make clustering labels noisy.The proposed framework targets the negative influence of these noise-prone pseudo-labels during adaptation.
- Uncertainty-guided optimization: Samples with wrong pseudo-labels generally have higher uncertainty, so UNRN assigns smaller credibility weights to their loss contributions.The paper uses ω_i = exp(−u_i) for stable training rather than only the inverse uncertainty.
- Uncertainty estimation: Uncertainty is computed from the inconsistency between student and teacher soft multilabels relative to target and source reference-person features.The soft multilabels are obtained by softmax-normalizing similarities to cluster centers and source classifier class centers, then compared with KL divergence.
- Uncertainty-guided optimization: Uncertainty re-weights the ID classification, triplet, and contrastive losses, reducing gradients from less credible samples and pairs.The contrastive objective also retains self-paced weighting to emphasize harder sample pairs.
- Uncertainty-guided optimization: A regularization loss discourages persistently large uncertainty while λ_tri, λ_ct, and λ_reg control objective contributions.This regularizer prevents the first three losses from being reduced continually through low credibility weights.
4 Experiments
Experiments evaluate UNRN on three person ReID datasets using mAP and CMC accuracies, comparing state-of-the-art performance, ablations, uncertainty designs, and memory-bank sizes. UNRN consistently improves adaptation results, while uncertainty re-weighting and combined reference centers provide key gains.
- 4.1 Datasets and Evaluation Metrics: UNRN is evaluated on DukeMTMC-reID, Market-1501, and MSMT17 using mAP and CMC Rank-1/5/10 accuracy.The experiments use three benchmark datasets and report mean average precision together with CMC accuracies.
- 4.3 Comparison with the State-of-the-arts: 2.1%, 3.0%, 2.4%, and 2.9% mAP gains over the second-best UDA methods are reported for Duke→Market, Market→Duke, Market→MSMT, and Duke→MSMT.These four domain-adaptation settings are compared in Table 1.
- 4.3 Comparison with the State-of-the-arts: UNRN uses one network with two models and still outperforms the best-performing MEB-Net, which uses three networks with six models.The comparison highlights a lower model count during training than several mutual mean-teacher approaches.
- 4.4 Ablation Studies: Applying uncertainty-guided optimization to ID, triplet, and contrastive losses improves mAP over SBase. by 2.7% for Duke→Market and 4.3% for Market→Duke.Replacing ID with UID improves mAP by 1.9% and 3.4%, while replacing triplet loss with UTRI yields similar improvements.
- 4.5 Design Choices: Soft multilabel consistency estimates uncertainty more effectively than feature consistency, and jointly using target and source centers performs best.Source centers alone are similar to SBase., target centers improve substantially, and the combined design is strongest.
- 4.5 Design Choices: UNRN-R exceeds the version without regularization by 0.6% mAP on Duke→Market and 1.1% on Market→Duke, while memory-bank performance saturates around N=8192.Increasing the queue length improves performance until the memory-bank size is approximately 8192.
5 Conclusion
The paper uses uncertainty from student–teacher consistency to assess pseudo-label reliability and re-weight ReID losses, reducing the influence of noisy labels during adaptation. The method improves over a strong baseline and reaches state-of-the-art performance on benchmark datasets.
- UNRN estimates target-sample pseudo-label reliability from the inconsistency between mean teacher and student outputs.The paper associates weaker student–teacher consistency with generally incorrect clustering pseudo-labels.
- Uncertainty re-weights ID classification, triplet, and contrastive losses to reduce noisy pseudo-label influence.
- The uncertainty-guided optimization brings significant improvement over the strong baseline.
- The proposed scheme achieves state-of-the-art performance on benchmark datasets.
7 Ethical Impact
Person re-identification can support applications such as smart retail, but identifying pedestrians may violate human privacy. The paper calls for strict regulations and laws to ensure legal use and protect data.
- The method can facilitate smart retail systems by matching or identifying people across images.
- Using person re-identification to identify pedestrians may violate human privacy.
- Governments and officials should formulate strict regulations and laws for legal ReID use and data protection.