[Paper Review] AD-Cluster: Augmented Discriminative Clustering for Domain Adaptive Person Re-identification

Yunpeng Zhai, Shijian Lu+5 more|arXiv (Cornell University)|Apr 19, 2020

Video Surveillance and Tracking MethodsComputer Science52 references32 citations

TL;DR

AD-Cluster proposes an unsupervised domain adaptation method for person re-ID by iteratively clustering target-domain samples, augmenting them across cameras with StarGAN, and learning discriminative features via a min-max adversarial framework.

ABSTRACT

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.

Motivation & Objective

• Motivate domain adaptive person re-ID when target identities are unknown.
• Leverage abundant unlabeled target-domain samples via clustering and augmentation.
• Improve discrimination by jointly learning sample diversity and feature separation.
• Demonstrate gains over state-of-the-art on Market-1501 and DukeMTMC-reID datasets.

Proposed method

• Density-based clustering to generate pseudo-labels in the target domain.
• Adaptive sample augmentation across cameras using StarGAN to increase target-domain diversity while preserving pseudo-labels.
• Adversarial min-max learning between an image generator and a feature encoder to increase sample diversity and tighten intra-cluster feature distances.
• Triplet-based discriminative learning to minimize intra-cluster distances and maximize inter-cluster distances.
• Iterative re-clustering and re-training to progressively refine pseudo-labels and representations.

Experimental results

Research questions

• RQ1 How can unlabeled target-domain samples be effectively exploited for domain adaptive re-ID without manual labeling?
• RQ2 Can cross-camera image generation preserve identity while enriching sample diversity to improve discriminative learning?
• RQ3 Does an adversarial min-max framework between a generator and a feature encoder improve cluster quality and re-ID performance?
• RQ4 What are the contributions of clustering, augmentation, and discriminative learning to transfer performance across DukeMTMC-reID and Market-1501?

Key findings

• AD-Cluster achieves state-of-the-art unsupervised domain adaptation performance on DukeMTMC-reID → Market-1501 and Market-1501 → DukeMTMC-reID transfers.
• For DukeMTMC-reID → Market-1501, AD-Cluster reaches 86.7% rank-1 and 68.3% mAP (surpassing prior best by notable margins).
• For Market-1501 → DukeMTMC-reID, AD-Cluster attains 72.6% rank-1 and 54.1% mAP (outperforming prior best).
• Adaptive sample augmentation and discriminative learning contribute sequentially to substantial gains over baselines and ablations.
• Iterative min-max clustering improves pseudo-label accuracy and overall re-ID metrics compared to density-based clustering alone.