Disentanglement-based Cross-Domain Feature Augmentation for Effective Unsupervised Domain Adaptive Person Re-identification

Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Quanzeng You, Zicheng Liu, Kecheng Zheng, Zhibo ChenarXiv 2021 · arXiv

Introduction

Person Re-identification (ReID) is a critical computer vision task focused on identifying the same individual across non-overlapping camera views. While supervised learning has achieved significant success, its performance often degrades when deployed in new environments due to the "domain gap"—variations in lighting, camera angles, and backgrounds. Unsupervised Domain Adaptation (UDA) seeks to bridge this gap by transferring knowledge from a labeled source domain to an unlabeled target domain.

Traditional UDA methods often rely on pseudo-labels generated via clustering. However, these labels are frequently noisy, and the lack of diverse target-domain data limits model robustness. To address this, researchers have introduced Disentanglement-based Cross-Domain Feature Augmentation (DCDFA), a novel strategy that synthesizes high-quality training data directly in the feature space.

Methodology: The DCDFA Framework

The core innovation of DCDFA is the disentanglement of person features into two distinct components: an Identity Base ($b$), which captures domain-invariant characteristics of the individual, and a Domain-Specific Enhancement ($e$), which encapsulates environmental styles such as illumination and camera quirks.

Feature Decomposition and Recomposition

As illustrated in the architecture below, the process involves three main stages:

  1. Extraction: A backbone network (e.g., ResNet-50) extracts high-level features from both source and target images.
  2. Decomposition: Using a channel-wise attention module, the extracted features are split into the identity base and the domain-specific enhancement. This allows the model to isolate "who" the person is from "where" they are.
  3. Recomposition (Feature Swapping): The model performs cross-domain augmentation by pairing the reliable identity base from a labeled source image with the environmental enhancement from a target image. This creates a synthetic feature that represents a known source identity as if it were captured in the target environment.
The DCDFA architecture showing the Extraction, Decomposition, and Recomposition stages for feature augmentation.

Optimization Objectives

To ensure effective disentanglement and realistic augmentation, the framework employs two primary loss functions:

  • Cross-Domain ReID Loss ($\\mathcal{L}_{RReID}$): Ensures that the recomposed features maintain the correct identity across domains.
  • Domain Classification Loss ($\\mathcal{L}_{Domain}$): Forces the enhancement component to accurately capture the statistical distribution of the target domain.

Experimental Setup and Datasets

The DCDFA method was evaluated on three benchmark datasets: Market-1501, DukeMTMC-reID, and MSMT17. These datasets provide a rigorous testing ground for cross-domain transfer, such as training on Market-1501 and adapting to the more complex DukeMTMC-reID environment.

Key Results and Discussion

DCDFA consistently outperforms existing state-of-the-art (SOTA) methods. By generating diverse and accurately labeled features for the target domain, it significantly improves the model's discriminative power.

As shown in the ablation study below, the addition of the Domain Classification Loss and the Cross-Domain ReID Loss to the strong baseline (SBase) leads to a substantial performance boost. Specifically, the full DCDFA model achieves 68.9% mAP on the Market $\to$ Duke transfer task and 78.6% mAP on Duke $\to$ Market.

Ablation study demonstrating the performance gains achieved by incorporating the DCDFA loss components.

Beyond accuracy, DCDFA is highly efficient. Unlike generative methods that use GANs to create synthetic images, feature-level augmentation adds minimal training overhead (less than 5%) and zero computational cost during inference, as the disentanglement modules are only used during the training phase.

Conclusion

Disentanglement-based Cross-Domain Feature Augmentation (DCDFA) offers a powerful and efficient solution for UDA in person ReID. By treating environmental style as an asset for data synthesis rather than noise to be ignored, DCDFA bridges the domain gap more effectively than traditional methods. Future work may explore combining this feature-level augmentation with advanced self-paced contrastive learning techniques to further enhance performance on massive, multi-scene datasets.

Summary updated

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.
BibTeX
@misc{zhang2021disentanglement,
  author = {Zhizheng Zhang and Cuiling Lan and Wenjun Zeng and Quanzeng You and Zicheng Liu and Kecheng Zheng and Zhibo Chen},
  title = {{Disentanglement-based Cross-Domain Feature Augmentation for Effective Unsupervised Domain Adaptive Person Re-identification}},
  year = {2021},
  eprint = {2103.13917},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.13917},
}
Figure