Source-linked AI summary
AD-Cluster: Augmented Discriminative Clustering for Domain Adaptive Person Re-identification
Yunpeng Zhai, Shijian Lu, Qixiang Ye, Xuebo Shan, Jie Chen, Rongrong Ji, Yonghong Tian
TL;DR
Domain adaptive person re-ID must generalize across domains with unknown target identities, but existing methods insufficiently exploit unlabeled target samples. AD-Cluster iteratively clusters target data, augments samples across cameras, and learns discriminative features through adversarial min-max optimization. It achieves state-of-the-art UDA accuracy on two large-scale benchmarks, while performance depends on generated-sample quality and target-domain camera-style variation.
Problem
Domain adaptive re-ID must learn from labeled source data and unlabeled target images whose identities are unavailable, while existing methods largely ignore target-domain unlabeled samples.
Method
AD-Cluster combines density-based pseudo-label clustering, adaptive cross-camera sample augmentation, and discriminative feature learning in an iterative adversarial min-max scheme.
Results
AD-Cluster produces state-of-the-art UDA accuracy on Market-1501 and DukeMTMC-reID and achieves significant gains over prior methods.
Takeaways & Limitations
Iteratively increasing target-domain sample diversity while learning discriminative features improves the discrimination ability of deep re-ID models.
Takeaways & Limitations
Performance depends on generated-sample quality and the strength of camera-style invariance in the target-domain sample distribution.
Abstract
from arXiv · showhide
Domain adaptive person re-identification (re-ID) is a challenging task, especially when person identities in target domains are unknown. Existing methods attempt to address this challenge by transferring image styles or aligning feature distributions across domains, whereas the rich unlabeled samples in target domains are not sufficiently exploited. This paper presents a novel augmented discriminative clustering (AD-Cluster) technique that estimates and augments person clusters in target domains and enforces the discrimination ability of re-ID models with the augmented clusters. AD-Cluster is trained by iterative density-based clustering, adaptive sample augmentation, and discriminative feature learning. It learns an image generator and a feature encoder which aim to maximize the intra-cluster diversity in the sample space and minimize the intra-cluster distance in the feature space in an adversarial min-max manner. Finally, AD-Cluster increases the diversity of sample clusters and improves the discrimination capability of re-ID models greatly. Extensive experiments over Market-1501 and DukeMTMC-reID show that AD-Cluster outperforms the state-of-the-art with large margins.
1. Introduction
Domain adaptive person re-ID remains difficult because models generalize poorly across changing camera and appearance conditions, while many UDA methods underuse unlabeled target samples. AD-Cluster addresses this gap by combining clustering, sample augmentation, and discriminative feature learning in an adversarial min-max framework.
- The generator maximizes intra-cluster sample distance, while the feature encoder minimizes intra-cluster feature distance in an adversarial min-max scheme.
- Domain adaptive re-ID remains an open challenge because camera configurations, lighting conditions, and person views change across domains.
- Many UDA methods align feature distributions or transfer source images into target styles but largely ignore unlabeled target samples.
- AD-Cluster predicts target-domain sample clusters using density-based clustering and alternates sample generation with discriminative feature learning.The generator preserves pseudo-identity labels while translating clustered images across cameras; the encoder separates clusters and contracts each cluster in feature space.
- AD-Cluster reports significant performance gains over the state of the art on Market-1501 and DukeMTMC-reID.
2. Related Works
Prior domain adaptive person re-ID work includes domain alignment, domain-invariant feature learning, generative style transfer, and pseudo-label prediction. AD-Cluster focuses on exploiting unlabeled target samples by iteratively augmenting clusters to improve representation of difficult samples.
- UDA approaches: UDA methods for person re-ID have pursued domain alignment, domain-invariant features, generative image transfer, and pseudo-label prediction.
- Domain alignment: Domain alignment methods reduce source-target distribution variation through feature statistics, attributes, camera-invariant similarities, or discrepancy minimization.
- GAN-based methods: GAN-based methods transfer identity-labeled source images into target-domain styles to train discriminative re-ID models.
- Remaining challenge: Existing methods often preserve source discriminative information but underuse unlabeled target samples, limiting their ability to maximize target-domain discrimination capacity.
- Pseudo-label prediction: Pseudo-label approaches cluster or match unlabeled target samples, while refinement methods incorporate camera invariance into label estimation.
- AD-Cluster: AD-Cluster addresses difficult samples by iteratively generating diverse, representative target-domain samples and including them in discriminative learning.
3. The Proposed Approach
AD-Cluster adapts person re-ID by clustering unlabeled target samples, augmenting them across camera styles, and jointly learning diverse yet discriminative representations.
- Target-domain clustering: AD-Cluster uses a source-trained encoder to extract target features, then applies density-based clustering and assigns cluster IDs as pseudo-labels.Each iteration extracts features, computes k-reciprocal distances, clusters samples, and assigns pseudo-labels.
- Source-domain pretraining: The source-domain encoder is initially trained with classification and ranking losses before target-domain clustering and adaptation.The source objective is Lsrc = Lcls + Ltri.
- Adaptive sample augmentation: Target-domain clustering addresses noisy pseudo-labels and limited within-cluster sample diversity, which hinder discriminative representation learning.The method augments target samples to increase diversity while preserving identities and adding camera, lighting, and view variation.
- Adaptive sample augmentation: StarGAN generates camera-style variants for clustered target images while preserving their pseudo-labels, increasing each cluster’s sample count by a factor of K −1.The generated images and originals are used for discriminative feature learning.
- Adversarial min-max optimization: AD-Cluster alternates generator and encoder updates: the generator maximizes intra-cluster sample diversity, while the encoder minimizes intra-cluster and maximizes inter-cluster feature distances.The encoder uses triplet loss, and the diversity loss is combined with StarGAN’s loss to optimize the generator.
- Adversarial min-max optimization: With generated and original target images, triplet learning uses same-cluster positives and hardest different-cluster negatives to strengthen target-domain discrimination.The objective selects all positive samples and the hardest negative sample for each anchor.
4. Experiments
Experiments evaluate AD-Cluster on Market1501 and DukeMTMC-ReID using CMC and mAP, comparing its clustering, augmentation, and discriminative-learning components with transfer baselines and state-of-the-art methods. AD-Cluster achieves the strongest reported transfer performance, while ablations and parameter analyses examine its components and behavior.
- Comparisons with State-of-the-Arts: AD-Cluster significantly outperforms all compared methods on both DukeMTMC-reID →Market1501 and Market1501 →DukeMTMC-reID transfers.Table 1 evaluates competing unsupervised and UDA-based re-ID methods using CMC and mAP.
- Comparisons with State-of-the-Arts: 86.7% rank-1 accuracy and 68.3% mAP are achieved for DukeMTMC-reID →Market1501, exceeding UDAP by 10.9% and 14.6%, respectively.For Market1501 →DukeMTMC-reID, the corresponding results are 72.6% rank-1 accuracy and 54.1% mAP, exceeding UDAP by 4.2% and 5.1%.
- Baseline, the Upper and Lower Bounds: 46.3% rank-1 accuracy for direct transfer rises to 73.8% with density-based clustering on Market1501 evaluation.Across the two evaluated datasets, density-based clustering improves rank-1 accuracy from 46.3% to 73.8% and from 28.0% to 68.6%.
- Adaptive Sample Augmentation: 83.3% rank-1 accuracy and 62.8% mAP from Baseline+ASA exceed the Baseline by 9.5% and 11.8% for DukeMTMC-reID →Market1501.The reported improvement is associated with ID-preserving cross-camera sample augmentation and improved feature-space sample distributions.
- Ablation Studies: 13% higher rank-1 accuracy and 17% higher mAP are obtained by complete AD-Cluster relative to the Baseline in ablation studies.Baseline+ASA+DL reaches 86.7% rank-1 accuracy for DukeMTMC-reID →Market1501, only 5.2% below the corresponding supervised model.
- Parameter Analysis: λ = 0.03 gives the best reported performance, while smaller λ generally increases cluster diversity but may weaken identity preservation.The min-max attenuation coefficient affects both mAP and rank-1 accuracy on Market-1501.
5. Conclusion
AD-Cluster combines density-based clustering, adaptive sample augmentation, and min-max optimization to improve domain-adaptive person re-ID. Experiments on two large-scale benchmarks establish new state-of-the-art UDA accuracy.
- AD-Cluster combines density-based clustering, adaptive sample augmentation, and min-max optimization to learn a more discriminative re-ID model.The method generates more diverse samples while improving the model’s discrimination ability.
- Experiments demonstrate that adaptive sample augmentation and min-max optimization improve the discrimination ability of deep re-ID models.
- AD-Cluster achieves new state-of-the-art UDA accuracy on Market-1501 and DukeMTMC-reID.
- The approach provides a fresh insight for broader unsupervised domain adaptation problems in recognition and detection.