[Paper Review] Disentanglement-based Cross-Domain Feature Augmentation for Effective Unsupervised Domain Adaptive Person Re-identification

Zhizheng Zhang, Cuiling Lan+5 more|arXiv (Cornell University)|Mar 25, 2021

Video Surveillance and Tracking Methods41 references4 citations

TL;DR

This paper proposes Disentanglement-based Cross-Domain Feature Augmentation (DCDFA), a novel unsupervised domain adaptation method for person re-identification that disentangles features into domain-invariant (identity-preserving) and domain-specific components, then performs cross-domain feature recomposition to generate diverse, reliable augmented training samples. DCDFA achieves state-of-the-art performance by enabling robust identity generalization and improved distribution approximation without increasing inference cost.

ABSTRACT

Unsupervised domain adaptive (UDA) person re-identification (ReID) aims to transfer the knowledge from the labeled source domain to the unlabeled target domain for person matching. One challenge is how to generate target domain samples with reliable labels for training. To address this problem, we propose a Disentanglement-based Cross-Domain Feature Augmentation (DCDFA) strategy, where the augmented features characterize well the target and source domain data distributions while inheriting reliable identity labels. Particularly, we disentangle each sample feature into a robust domain-invariant/shared feature and a domain-specific feature, and perform cross-domain feature recomposition to enhance the diversity of samples used in the training, with the constraints of cross-domain ReID loss and domain classification loss. Each recomposed feature, obtained based on the domain-invariant feature (which enables a reliable inheritance of identity) and an enhancement from a domain specific feature (which enables the approximation of real distributions), is thus an "ideal" augmentation. Extensive experimental results demonstrate the effectiveness of our method, which achieves the state-of-the-art performance.

Motivation & Objective

• To address the challenge of unreliable pseudo-labels in unsupervised domain adaptive (UDA) person re-identification due to domain shift between source and target domains.
• To generate high-quality, diverse training samples for the target domain that approximate real data distributions while preserving reliable identity labels.
• To improve feature generalization and robustness in UDA ReID by leveraging disentangled representations and cross-domain feature recomposition.
• To provide a generic, efficient augmentation strategy compatible with existing UDA methods to further boost performance.

Proposed method

• Disentangles each input feature into a domain-invariant/shared feature (preserving identity) and a domain-specific feature (encoding domain characteristics).
• Performs cross-domain feature recomposition by combining a domain-invariant feature from one domain with a domain-specific feature from the other domain to generate augmented samples.
• Optimizes the disentanglement and augmentation using a combined loss: cross-domain ReID loss for identity consistency and domain classification loss for domain-specific feature learning.
• Employs a learnable attention module to guide channel-wise feature disentanglement, with ablation showing channel-wise attention as most effective.
• Introduces a regularization loss on the identity base feature to enhance discriminativeness and stabilize training.
• Maintains inference efficiency by discarding the disentanglement module during testing, adding less than 5% training time overhead.

Experimental results

Research questions

• RQ1 Can disentangled feature decomposition improve the reliability of pseudo-labeled samples in unsupervised domain adaptation for person re-identification?
• RQ2 Can cross-domain feature recomposition generate augmented samples that better approximate real data distributions while preserving identity labels?
• RQ3 Does combining domain-invariant and domain-specific features through disentanglement lead to improved generalization and performance in UDA ReID?
• RQ4 How does the proposed DCDFA strategy compare in efficiency and performance to existing SOTA methods in UDA ReID?

Key findings

• DCDFA achieves state-of-the-art performance on Market-1501 → DukeMTMC-reID and DukeMTMC-reID → Market-1501 settings, with mAP improvements of 2.7% and 1.8% respectively over the baseline.
• On Market-1501 → MSMT17, DCDFA achieves 27.1% mAP, slightly below SpCL (28.0%), but remains highly competitive and complementary to self-paced clustering methods.
• Joint optimization of the identity disentanglement loss (ℒCID) and domain classification loss (ℒDomain) leads to significant performance gains, demonstrating the importance of explicit supervision on both feature components.
• The ablation study confirms that channel-wise attention is the most effective mechanism for feature disentanglement, outperforming spatial and channel-spatial attention.
• The method introduces less than 5% additional training time and incurs no inference cost, making it highly efficient and deployable.
• DCDFA is generic and effective as a plug-in augmentation strategy, significantly boosting performance when applied on top of strong baselines like SBase.