Mask-Guided Region Attention Network for Person Re-Identification

Cong Zhou, Han YuLecture Notes in Computer Science 2020 · DOI · OpenAlex

Abstract

Person Re-Identification (ReID) remains a challenging task in computer vision due to background clutter, varying camera viewpoints, and the dynamic nature of human gait, which often leads to self-occlusion. The Mask-Guided Region Attention Network (MGRAN) addresses these issues by moving beyond traditional, imprecise bounding-box alignments. By utilizing pixel-level pedestrian masks and a novel region-level alignment strategy (Head, Upper Body, and Lower Body), MGRAN effectively eliminates background noise and provides a stable feature representation for identifying individuals across non-overlapping camera views.

Introduction

Identifying the same individual across different surveillance cameras is a cornerstone of modern security and urban analytics. However, standard ReID methods often struggle with "self-occlusion"—where a person's limbs hide each other during movement—and background noise that can confuse feature extraction. While many state-of-the-art (SOTA) methods attempt to use human pose estimation to align body parts, they frequently rely on rectangular Regions of Interest (RoIs) that include irrelevant background pixels or attempt to match highly mobile limbs (like arms and legs) that are prone to misalignment.

The MGRAN Framework

The proposed MGRAN framework introduces two primary modules designed to refine feature extraction and ensure morphological stability: the Mask-Guided Region Attention (MRA) module and the Multi-feature Alignment (MA) module.

Mask-Guided Region Attention (MRA)

Instead of relying on simple bounding boxes, MGRAN utilizes a pre-trained Mask R-CNN to generate precise binary pedestrian masks. This allows the network to focus exclusively on the pedestrian's silhouette at a pixel level, effectively "masking out" background clutter such as sidewalks, vehicles, or other pedestrians.

The MGRAN pipeline: integrating binary masks and keypoints to isolate the pedestrian from background noise.

Multi-feature Alignment (MA)

The MA module processes the masked image through two distinct stages:

  1. Space Alignment (SA): The original image is multiplied by the binary mask (Hadamard product) to strip away background noise. Using four detected keypoints, the system then divides the pedestrian into three stable regions: the Head, Upper Body, and Lower Body.
  2. Multi-feature Fusion (MF): To capture both fine-grained details and the overall silhouette, the framework employs four parallel ResNet-50 branches. These branches extract features from the global pedestrian region and the three local sub-regions, which are then concatenated into a final robust representation.

Region-Level vs. Part-Level Alignment

A key insight of MGRAN is the shift from "part-level" to "region-level" alignment. The authors argue that matching specific limbs (arms, legs) is inherently unreliable because their positions change drastically during a gait cycle. In contrast, broader regions—Head, Upper Body, and Lower Body—remain relatively stable and are less affected by self-occlusion.

Region-level alignment: dividing the pedestrian into stable morphological segments (Head, Upper Body, Lower Body).

By focusing on these stable regions, the model avoids the "misalignment" errors common in pose-based models that try to track individual joints. Furthermore, the inclusion of a "Global" feature branch ensures that the overall body shape and posture are preserved, providing a holistic context for the local region features.

Experimental Results

The authors evaluated MGRAN on three standard large-scale ReID benchmarks, where it consistently outperformed existing pose-relevant and pose-irrelevant SOTA methods:

  • Market-1501: Achieved 88.91% Rank-1 accuracy and 78.03% mAP.
  • DukeMTMC-reID: Reached 78.12% Rank-1 accuracy.
  • CUHK03: On the "Labeled" dataset, it achieved 93.02% Rank-1, a significant improvement over previous benchmarks.

Ablation studies confirmed that both the pixel-level masking and the multi-feature fusion were critical to these results. Removing the global feature branch or reverting to traditional part-level alignment led to a measurable decrease in accuracy.

Conclusion and Future Work

MGRAN demonstrates that pixel-level precision and stable region-based alignment are superior to traditional bounding-box and part-based methods for person ReID. While the use of four ResNet-50 branches increases computational overhead, the gains in accuracy and robustness against background noise are substantial. Future research may focus on optimizing the model's efficiency and exploring the use of Generative Adversarial Networks (GANs) to augment training data for even more challenging, low-light, or crowded environments.

Summary updated

BibTeX
@inproceedings{zhou2020mask,
  author = {Cong Zhou and Han Yu},
  title = {{Mask-Guided Region Attention Network for Person Re-Identification}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2020},
  doi = {10.1007/978-3-030-47436-2\_22},
}
Figure