[Paper Review] Attributes-aided Part Detection and Refinement for Person Re-identification

Shuzhao Li, Huimin Yu+2 more|arXiv (Cornell University)|Feb 27, 2019

Video Surveillance and Tracking Methods20 references4 citations

TL;DR

This paper proposes APDR, a novel deep learning model that leverages attribute detection as pose-invariant part detectors to address body part misalignment in person re-identification. By fusing attribute information to refine local features, APDR achieves state-of-the-art performance on Market-1501 and DukeMTMC-reID, outperforming existing methods in both rank-1 accuracy and mAP, especially after re-ranking.

ABSTRACT

Person attributes are often exploited as mid-level human semantic information to help promote the performance of person re-identification task. In this paper, unlike most existing methods simply taking attribute learning as a classification problem, we perform it in a different way with the motivation that attributes are related to specific local regions, which refers to the perceptual ability of attributes. We utilize the process of attribute detection to generate corresponding attribute-part detectors, whose invariance to many influences like poses and camera views can be guaranteed. With detected local part regions, our model extracts local features to handle the body part misalignment problem, which is another major challenge for person re-identification. The local descriptors are further refined by fused attribute information to eliminate interferences caused by detection deviation. Extensive experiments on two popular benchmarks with attribute annotations demonstrate the effectiveness of our model and competitive performance compared with state-of-the-art algorithms.

Motivation & Objective

• Address the persistent challenge of body part misalignment in person re-identification due to pose, viewpoint, and occlusion variations.
• Overcome the limitations of existing part detection methods that rely on pose estimation or attention models with weak supervision.
• Leverage the perceptual and semantic consistency of human attributes to guide more reliable and discriminative local part localization.
• Improve feature representation by fusing attribute information to refine local descriptors and reduce noise from detection errors.
• Develop a unified framework that integrates attribute learning, part detection, and feature refinement for end-to-end re-identification optimization.

Proposed method

• Train an attribute detection head to generate part detectors that are invariant to pose and viewpoint by learning local region associations with attributes.
• Use detected attribute-part regions as part localizers to extract localized deep features, reducing spatial misalignment issues.
• Introduce a simple attribute fusion module to combine multiple attributes into a compact, semantically meaningful representation.
• Refine the extracted local features using the fused attribute representation to suppress irrelevant or noisy features caused by detection deviation.
• Combine the refined local features with global image-level features for final re-identification scoring.
• Apply re-ranking as a post-processing step to further boost performance, leveraging the model's robustness and discriminative capacity.

Experimental results

Research questions

• RQ1 Can attribute detection be effectively repurposed as a pose-invariant part detection mechanism for person re-identification?
• RQ2 How does fusing multiple attributes improve the reliability and discriminativeness of local part features compared to individual attribute classification?
• RQ3 To what extent can attribute-aided part detection reduce the impact of detection errors and spatial misalignment in re-identification?
• RQ4 Does the proposed refinement mechanism based on fused attribute information lead to more robust local feature representations than standard part pooling or attention mechanisms?
• RQ5 How does the model perform under challenging conditions such as large pose variations and complex backgrounds on large-scale benchmarks?

Key findings

• APDR achieves 93.1% rank-1 accuracy and 80.1% mAP on Market-1501 without re-ranking, outperforming most state-of-the-art methods including PABR and RPP.
• After applying re-ranking, APDR achieves 94.4% rank-1 and 90.2% mAP on Market-1501, surpassing even the RPP model and setting a new SOTA performance.
• On DukeMTMC-reID, APDR achieves 84.3% rank-1 and 69.7% mAP, matching or exceeding the performance of PABR and RPP despite not using dilation structures.
• With re-ranking, APDR reaches 87.3% rank-1 and 83.2% mAP on DukeMTMC-reID, demonstrating strong generalization across diverse surveillance conditions.
• The model significantly outperforms attribute-based methods like APR and ACRN, proving that perception-driven part detection via attributes is more effective than simple attribute classification.
• The ablation shows that attribute fusion and refinement are critical, as the model’s performance degrades when these components are removed, confirming their functional importance.