LRDNN: Local-refining based Deep Neural Network for Person Re-Identification with Attribute Discerning

Qinqin Zhou, Bineng Zhong, Xiangyuan Lan, Gan Sun, Yulun Zhang, Mengran Gou

Abstract

Person Re-Identification (re-ID) remains a challenging task in computer vision due to significant variations in human pose, camera viewpoints, and environmental clutter. This paper introduces the Local-refining based Deep Neural Network (LRDNN), a multi-task framework designed to enhance re-ID performance by integrating identity recognition, body attribute discerning, and skeletal pose estimation. A key innovation, the Channel Parse Block (CPB), is employed to robustly filter noise from unstable pose estimations. LRDNN achieves state-of-the-art results on the Market1501 and DukeMTMC-reid benchmarks.

Introduction

The goal of person re-identification is to identify the same individual across multiple non-overlapping camera views. This task is fundamental for large-scale surveillance and forensic applications. However, three primary obstacles hinder performance:

  1. Pose Variation: Individuals appear differently depending on their orientation relative to the camera.
  2. Misalignment: Varying camera angles and heights cause body parts to appear in different spatial locations within the frame.
  3. Background Clutter: Complex environments introduce noise that can contaminate global feature representations.

While auxiliary information like body attributes and pose estimation can provide valuable cues, these modules are often unreliable. Inaccurate pose estimation (e.g., misidentifying a shoulder as a knee) can degrade the overall system. LRDNN addresses this by robustly integrating these local cues to refine global identification features.

Methodology

The LRDNN architecture is a dual-branch system that fuses global appearance with local spatial information.

Architecture Overview

  • Main Branch: Utilizing a ResNet-50 backbone, this branch extracts high-level semantic features. It is trained multi-task style to identify individuals while simultaneously discerning specific attributes such as gender, hair length, and clothing type.
  • Pose Branch: This branch employs a multi-stage CNN (inspired by OpenPose) to generate part confidence maps and part affinity fields, identifying the locations of joints and limbs.
  • Channel Parse Block (CPB): To mitigate the instability of pose estimation, the CPB acts as a dynamic filter. It learns to weight feature channels in the pose branch, prioritizing reliable body parts and suppressing noisy or irrelevant data.
  • Compact Bilinear Pooling (CBP): The features from both branches are fused using CBP. This method captures complex interactions between pose and appearance data efficiently, refining the global feature with precise local spatial knowledge.

Attribute Discerning

By requiring the network to predict attributes alongside identity, the model learns more granular and discriminative features. Attributes like "carrying a backpack" or "long hair" serve as critical anchors for distinguishing between individuals with similar overall appearances. This "attribute discerning" capability allows the model to use local details as tie-breakers in ambiguous cases.

Examples of attribute discerning in LRDNN, highlighting features like bags, backpacks, hair length, and hats.

Experimental Results

The authors evaluated LRDNN on two major benchmarks: Market1501 and DukeMTMC-reid. The pose branch was pre-trained on the MSCOCO dataset to establish a strong anatomical baseline.

  • Market1501: Achieved a Rank-1 accuracy of 90.4% and a Mean Average Precision (mAP) of 82.8%.
  • DukeMTMC-reid: Achieved a Rank-1 accuracy of 85.3% and an mAP of 73.2%.

Ablation studies confirmed the efficacy of the multi-task approach. Integrating attribute discerning alone improved the baseline by nearly 7%. The addition of the pose branch and the CPB provided the final performance boost, demonstrating that local refinement significantly outperforms purely global feature extraction.

Conclusion

LRDNN presents a robust framework for person re-identification by effectively leveraging local cues. The introduction of the Channel Parse Block ensures that inaccurate pose data does not negatively impact the system, while attribute discerning provides granular features for identity verification. Future work may focus on optimizing computational efficiency for real-time applications and enhancing the model's ability to handle extreme occlusions.

Summary updated

Abstract

Recently, pose or attribute information has been widely used to solve person re-identification (re-ID) problem. However, the inaccurate output from pose or attribute modules will impair the final person re-ID performance. Since re-ID, pose estimation and attribute recognition are all based on the person appearance information, we propose a Local-refining based Deep Neural Network (LRDNN) to aggregate pose estimation and attribute recognition to improve the re-ID performance. To this end, we add a pose branch to extract the local spatial information and optimize the whole network on both person identity and attribute objectives. To diminish the negative affect from unstable pose estimation, a novel structure called channel parse block (CPB) is introduced to learn weights on different feature channels in pose branch. Then two branches are combined with compact bilinear pooling. Experimental results on Market1501 and DukeMTMC-reid datasets illustrate the effectiveness of the proposed method.
BibTeX
@inproceedings{zhou2019lrdnn,
  author = {Qinqin Zhou and Bineng Zhong and Xiangyuan Lan and Gan Sun and Yulun Zhang and Mengran Gou},
  title = {{LRDNN: Local-refining based Deep Neural Network for Person Re-Identification with Attribute Discerning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/146},
}
Figure