Source-linked AI summary
Group-aware Label Transfer for Domain Adaptive Person Re-identification
Kecheng Zheng, Wu Liu, Lingxiao He, Tao Mei, Jiebo Luo, Zheng-Jun Zha
TL;DR
UDA-ReID must adapt across domain gaps without target annotations, while noisy pseudo labels can hinder feature learning. GLT combines online pseudo-label refinement with group-aware clustering and reports results that narrow the supervised–unsupervised gap.
Problem
Domain gaps make source-trained ReID models generalize poorly to unlabeled target domains, while noisy clustering labels can mislead feature learning.
Method
GLT jointly trains with pseudo labels and refines them online through optimal transport, while group-aware clustering generates multi-group pseudo labels.
Results
92.2 Rank-1 and 79.5 mAP are reported for Duke-to-Market, versus 94.1 Rank-1 and 85.7 mAP for supervised learning.
Takeaways & Limitations
GLT reports state-of-the-art performance across four UDA-ReID transfer tasks and significantly closes the gap with supervised performance.
Takeaways & Limitations
The transport formulation treats λ as a hyperparameter because final clustering and representation performance matter more than solving the optimal transport problem exactly.
Abstract
from arXiv · showhide
Unsupervised Domain Adaptive (UDA) person re-identification (ReID) aims at adapting the model trained on a labeled source-domain dataset to a target-domain dataset without any further annotations. Most successful UDA-ReID approaches combine clustering-based pseudo-label prediction with representation learning and perform the two steps in an alternating fashion. However, offline interaction between these two steps may allow noisy pseudo labels to substantially hinder the capability of the model. In this paper, we propose a Group-aware Label Transfer (GLT) algorithm, which enables the online interaction and mutual promotion of pseudo-label prediction and representation learning. Specifically, a label transfer algorithm simultaneously uses pseudo labels to train the data while refining the pseudo labels as an online clustering algorithm. It treats the online label refinery problem as an optimal transport problem, which explores the minimum cost for assigning M samples to N pseudo labels. More importantly, we introduce a group-aware strategy to assign implicit attribute group IDs to samples. The combination of the online label refining algorithm and the group-aware strategy can better correct the noisy pseudo label in an online fashion and narrow down the search space of the target identity. The effectiveness of the proposed GLT is demonstrated by the experimental results (Rank-1 accuracy) for Market1501$\to$DukeMTMC (82.0\%) and DukeMTMC$\to$Market1501 (92.2\%), remarkably closing the gap between unsupervised and supervised performance on person re-identification.
1. Introduction
UDA-ReID must adapt across domain gaps without target annotations, but noisy offline pseudo-label refinement can hinder representation learning. GLT addresses this through online group-aware label transfer and reports strong cross-domain results.
- UDA-ReID adapts models trained on annotated source data to unlabeled target data, where domain gaps undermine cross-scenario generalization.
- Clustering-based pseudo labels can be incorrect, mislead feature learning, and require time-consuming offline refinement over the entire dataset.
- Its group-aware strategy assigns implicit attribute groups and multi-group pseudo labels to narrow the target identity search space without manual grouping.
- GLT jointly integrates pseudo-label prediction and feature learning by using pseudo labels for training while refining them online.
- 92.2 Rank-1 and 79.5 mAP are reported for Duke-to-Market, compared with 94.1 Rank-1 and 85.7 mAP for supervised learning.
- GLT reports significant improvements over state-of-the-art approaches on Market-to-Duke, Duke-to-Market, Market-to-MSMT, and Duke-to-MSMT tasks.
2. Related Work
Related UDA-ReID work commonly relies on clustering, mutual learning, domain translation, or memory banks. These approaches address pseudo-label noise or cross-domain representation learning with differing complexity and computational requirements.
- Clustering-based methods generate hard or soft pseudo labels and train models using those labels because of their strong performance.
- Mutual-learning methods use collaborative networks and soft pseudo labels to suppress noise, while NRMT adds mutual instance selection.
- NRMT's hard reliable-instance selection requires threshold tuning and may discard useful information from abandoned samples.
- Domain-translation methods use translated source-to-target images and source identities, but GAN-based training may add computational cost and lack real-world applicability.
- Memory-bank methods exploit global instance similarity and enable hard-negative mining or invariance constraints across training data.
3. Group-aware Label Transfer
The framework combines clustering-based pseudo-label prediction with feature learning, then refines labels online through label transfer. Group-aware prototypes and optimal transport support minibatch-level pseudo-label refinement while avoiding trivial assignments.
- Clustering-based Method Revisit: UDA-ReID first pre-trains a model on labeled source data, extracts target features, and clusters target images into pseudo-label classes.These pseudo labels may contain noise because of the source–target domain gap and imperfect clustering.
- Clustering-based Method Revisit: Noisy pseudo labels can mislead feature learning, motivating online refinement during training rather than offline refinement over the entire dataset.The framework alternates pseudo-label generation and feature learning while correcting incorrect samples online.
- Group-aware Label Transfer: Each image feature is mapped to trainable prototype vectors to produce group probabilities, and the pseudo-label matrix supervises these probabilities during label transfer.The prototypes may be implemented with non-parametric or parametric classifiers and updated jointly with the image encoder.
- Group-aware Label Transfer: Equipartition constraints assign each sample to one label while distributing samples across classes, preventing the trivial solution in which all images share one pseudo label.The constraint is adapted to minibatches or the whole dataset.
- Group-aware Label Transfer: The label-refining assignment is formulated as optimal transport, minimizing assignment cost while constraining row and column marginals to preserve a source-domain-like class distribution.The transport matrix represents refined conditional pseudo-label probabilities for samples and labels.
- Group-aware Label Transfer: Sinkhorn-Knopp solves the large-scale transport problem through entropy regularization and iterative matrix scaling using feature and prototype matrices.The resulting probability matrix Q refines pseudo labels online; λ controls mapping smoothness and is kept high in practice to avoid collapse.
4. Experiments
Experiments evaluate GLT on three person ReID datasets across four unsupervised domain adaptation settings, comparing it with state-of-the-art methods and its component variants. GLT improves over competing approaches and ablations support the contributions of label transfer, memory, and group-aware refinement.
- Comparison with State-of-the-art Methods: 21.2%, 15.8%, 13.3%, and 14.4% mAP improvements over SSG are reported on the four tasks.The comparison uses simpler network architectures and no extra local-feature computation.
- Comparison with State-of-the-art Methods: 8.3%, 4.1%, 3.6%, and 4.4% mAP improvements over MMT are reported across the four tasks using one model and no attention mechanism.GLT also exceeds MEB-Net by 3.5% and 3.1% mAP on Duke→Market and Market→Duke, respectively.
- Comparison with State-of-the-art Methods: 92.2% versus 94.1% top-1 accuracy is reported for unsupervised versus supervised Duke→Market transfer.The result is presented as narrowing the supervised–unsupervised performance gap without manual annotations.
- Ablation Studies: 7.8% rank-1 accuracy and 6.9% mAP gains follow introduction of label transfer on Duke→Market.The target memory bank raises rank-1 accuracy from 88.1% to 89.0% and mAP from 66.4% to 68.2%, while group-aware label transfer adds 11.3% rank-1 accuracy and 3.0% mAP.
- Ablation Studies: Table 3 evaluates triplet loss and weighted contrastive loss as alternative loss components.The supplied passages identify the comparison and its “w/o WCL” notation but do not report the table’s numerical outcomes.
5. Conclusion
The conclusion presents GLT as a unified framework combining pseudo-label prediction and ReID representation learning through online label refinement and group-aware assignment. It reports state-of-the-art performance and a substantially narrowed supervised–unsupervised gap.
- GLT combines accurate pseudo-label prediction and effective ReID representation learning in one unified optimization objective.Its holistic and immediate interaction between the two steps is intended to support UDA person ReID.
- Online label refinement is formulated as an optimal transport problem assigning M samples to N pseudo labels at minimal cost.The approach uses group-aware assignment of implicit attribute group IDs to samples.
- The online refinement and group-aware strategy correct noisy pseudo labels and narrow the target-identity search space.The conclusion attributes these functions to the combination of the two design elements.
- GLT achieves state-of-the-art performance while significantly closing the gap between supervised and unsupervised person ReID.