Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Attention Disturbance and Dual-Path Constraint Network for Occluded Person Re-identification

Jiaer Xia, Lei Tan, Pingyang Dai, Mingbo Zhao, Yongjian Wu, Liujuan Cao

TL;DR

This paper tackles occluded person re-identification by addressing the gap between synthetic background occlusion and real-world occlusion. It introduces Attention Disturbance Mask (ADM), an attack-based augmentation that disturbs attention with realistic occlusion-like noise, and a Dual-Path Constraint Module (DPC) that jointly optimizes holistic and occluded paths with shared parameters. The resulting ADP framework improves attention robustness and leverages holistic supervision to achieve state-of-the-art results on occluded benchmarks while remaining efficient at inference. Overall, ADM and DPC enhance generalization to diverse occlusions and can be integrated with existing transformer-based Re-ID systems to boost performance.

Abstract

Occluded person re-identification (Re-ID) aims to address the potential occlusion problem when matching occluded or holistic pedestrians from different camera views. Many methods use the background as artificial occlusion and rely on attention networks to exclude noisy interference. However, the significant discrepancy between simple background occlusion and realistic occlusion can negatively impact the generalization of the network. To address this issue, we propose a novel transformer-based Attention Disturbance and Dual-Path Constraint Network (ADP) to enhance the generalization of attention networks. Firstly, to imitate real-world obstacles, we introduce an Attention Disturbance Mask (ADM) module that generates an offensive noise, which can distract attention like a realistic occluder, as a more complex form of occlusion. Secondly, to fully exploit these complex occluded images, we develop a Dual-Path Constraint Module (DPC) that can obtain preferable supervision information from holistic images through dual-path interaction. With our proposed method, the network can effectively circumvent a wide variety of occlusions using the basic ViT baseline. Comprehensive experimental evaluations conducted on person re-ID benchmarks demonstrate the superiority of ADP over state-of-the-art methods.

Attention Disturbance and Dual-Path Constraint Network for Occluded Person Re-identification

TL;DR

This paper tackles occluded person re-identification by addressing the gap between synthetic background occlusion and real-world occlusion. It introduces Attention Disturbance Mask (ADM), an attack-based augmentation that disturbs attention with realistic occlusion-like noise, and a Dual-Path Constraint Module (DPC) that jointly optimizes holistic and occluded paths with shared parameters. The resulting ADP framework improves attention robustness and leverages holistic supervision to achieve state-of-the-art results on occluded benchmarks while remaining efficient at inference. Overall, ADM and DPC enhance generalization to diverse occlusions and can be integrated with existing transformer-based Re-ID systems to boost performance.

Abstract

Occluded person re-identification (Re-ID) aims to address the potential occlusion problem when matching occluded or holistic pedestrians from different camera views. Many methods use the background as artificial occlusion and rely on attention networks to exclude noisy interference. However, the significant discrepancy between simple background occlusion and realistic occlusion can negatively impact the generalization of the network. To address this issue, we propose a novel transformer-based Attention Disturbance and Dual-Path Constraint Network (ADP) to enhance the generalization of attention networks. Firstly, to imitate real-world obstacles, we introduce an Attention Disturbance Mask (ADM) module that generates an offensive noise, which can distract attention like a realistic occluder, as a more complex form of occlusion. Secondly, to fully exploit these complex occluded images, we develop a Dual-Path Constraint Module (DPC) that can obtain preferable supervision information from holistic images through dual-path interaction. With our proposed method, the network can effectively circumvent a wide variety of occlusions using the basic ViT baseline. Comprehensive experimental evaluations conducted on person re-ID benchmarks demonstrate the superiority of ADP over state-of-the-art methods.
Paper Structure (13 sections, 12 equations, 4 figures, 5 tables)

This paper contains 13 sections, 12 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Overview
  5. Attention Disturbance Mask
  6. Dual-Path Constraint Module
  7. Experiment
  8. Datasets and Evaluation Setting
  9. Comparison with State-of-the-art Methods
  10. Ablation Studies
  11. Discussions
  12. Visualization of the Attention
  13. Conclusion

Figures (4)

  • Figure 1: Visualization of attention to baseline and proposed ADP. (a) The baseline trained with the assistance of background occlusion failed to avoid the realistic occlusion in the testing set. (b) The ADP trained by the proposed reality-similar occlusion ADM performs well on both artificial and real occlusion.
  • Figure 2: The overview of the proposed Attention Disturbance and Dual-Path Constraint Network (ADP). To create the corresponding occlusion images, the transformed background patch is used as the carrier of the Attention Disturbance Mask (ADM) and covers a random region of the original image. Then the Dual-Path Constraint Module (DPC) simultaneously deals with holistic and occluded images. In Multi-Head Attention (MHA) stage, ADM maximizes the similarity between class token and masked patches to optimize the mask.
  • Figure 3: Visualization of the feature distribution on Occlude-Duke dataset. Circles denote the features of images while the colors represent different identities. (a) Baseline refers to the model trained on the images without extra occlusions. (b) The middle plot shows the results of model trained on the images occluded by the background. (c) Compared to the other two models, the model trained on our ADM can avoid the influence of obstacles well.
  • Figure 4: Visualization of attention maps on occlusion testing set in Occluded-Duke dataset. (a) occluded person images. (b) attention maps of baseline. (c) attention maps of ADP.