EANet: Enhancing Alignment for Cross-Domain Person Re-identification (1812.11369v1)

Abstract: Person re-identification (ReID) has achieved significant improvement under the single-domain setting. However, directly exploiting a model to new domains is always faced with huge performance drop, and adapting the model to new domains without target-domain identity labels is still challenging. In this paper, we address cross-domain ReID and make contributions for both model generalization and adaptation. First, we propose Part Aligned Pooling (PAP) that brings significant improvement for cross-domain testing. Second, we design a Part Segmentation (PS) constraint over ReID feature to enhance alignment and improve model generalization. Finally, we show that applying our PS constraint to unlabeled target domain images serves as effective domain adaptation. We conduct extensive experiments between three large datasets, Market1501, CUHK03 and DukeMTMC-reID. Our model achieves state-of-the-art performance under both source-domain and cross-domain settings. For completeness, we also demonstrate the complementarity of our model to existing domain adaptation methods. The code is available at https://github.com/huanghoujing/EANet.

Enhancing Alignment for Cross-Domain Person Re-identification: An Expert Overview

The paper presents EANet, a novel approach to address the challenges in cross-domain Person Re-identification (ReID). The authors propose techniques to enhance alignment in ReID models, aiming to improve both generalization and adaptation across domains. The research is significant in the context of decreasing performance when applying single-domain trained models to new, unseen domains, which requires domain adaptation strategies without target-domain identity labels.

Contributions and Methodology

EANet introduces three primary contributions:

1. Part Aligned Pooling (PAP): This method seeks to rectify alignment problems inherent in traditional pooling strategies like PCB, which divide feature maps into uniform horizontal stripes regardless of body part positioning. PAP pools features from keypoint delimited regions, enhancing alignment by using pose estimation to guide part definitions.
2. Part Segmentation (PS) Constraint: This innovation adds a segmentation module to ReID feature maps, supervised by pseudo segmentation labels, to improve feature localization and reduce redundancy. The PS constraint enhances model generalization capabilities by maintaining part-awareness in feature representation.
3. Integration for Domain Adaptation: By applying the PS constraint to unlabeled images from the target domain, EANet adapts effectively, demonstrating improved performance without the need for target-domain labels.

Experimental Validation

The paper validates EANet's efficacy through extensive experiments on datasets Market1501, CUHK03, and DukeMTMC-reID, achieving superior performance across both single-domain and cross-domain settings.

• Numerical Results: EANet sets a new state-of-the-art on cross-domain tasks, significantly outperforming existing methods. The improvements are marked by a notable increase in Rank-1 accuracy and mAP scores, particularly when PS constraints are used on both source and target domains.
• Comparison with Existing Models: EANet demonstrates complementarity to existing methods like style transfer (SPGAN) and label estimation, further amplifying results when combined.

Theoretical and Practical Implications

This research provides insights into the critical role of alignment in cross-domain ReID. The techniques developed could inform future research directions focusing on enhanced feature localization and adaptive segmentation strategies. Practically, EANet's architecture can be integrated into surveillance systems where model adaptation across domains with varying characteristics is imperative.

Future Directions

The authors outline possible future research directions, including an exploration of segmentation map-based regional pooling and the refinement of PS constraints to suit occlusion scenarios. Additionally, leveraging publicly available part segmentation datasets like COCO could further enhance the effectiveness of ReID models by providing precise part annotations.

In conclusion, this paper makes substantial contributions to the domain of person ReID by addressing cross-domain adaptation challenges with novel alignment strategies. The advancements in model generalization and domain adaptation introduced by EANet demonstrate promising results and open avenues for future innovations in the field.