High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification (2003.08177v4)

Abstract: Occluded person re-identification (ReID) aims to match occluded person images to holistic ones across dis-joint cameras. In this paper, we propose a novel framework by learning high-order relation and topology information for discriminative features and robust alignment. At first, we use a CNN backbone and a key-points estimation model to extract semantic local features. Even so, occluded images still suffer from occlusion and outliers. Then, we view the local features of an image as nodes of a graph and propose an adaptive direction graph convolutional (ADGC)layer to pass relation information between nodes. The proposed ADGC layer can automatically suppress the message-passing of meaningless features by dynamically learning di-rection and degree of linkage. When aligning two groups of local features from two images, we view it as a graph matching problem and propose a cross-graph embedded-alignment (CGEA) layer to jointly learn and embed topology information to local features, and straightly predict similarity score. The proposed CGEA layer not only take full use of alignment learned by graph matching but also re-place sensitive one-to-one matching with a robust soft one. Finally, extensive experiments on occluded, partial, and holistic ReID tasks show the effectiveness of our proposed method. Specifically, our framework significantly outperforms state-of-the-art by6.5%mAP scores on Occluded-Duke dataset.

• The paper presents a robust graph-based framework using high-order relation and topology modules to enhance occluded person re-identification.
• It introduces innovative components like an adaptive directional graph convolution and cross-graph embedded alignment for precise feature matching.
• Empirical evaluations on Occluded-Duke demonstrated a 6.5% increase in mAP, underscoring its potential for real-world surveillance applications.

High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification

This paper presents a robust framework for handling the increasingly important task of occluded person re-identification (ReID). The authors propose a method that draws upon advanced graph-based techniques to improve feature discrimination and alignment in the presence of occlusions. The solution leverages high-order information concerning both relational and topological aspects of data, significantly enhancing model performance on occluded datasets compared to current state-of-the-art methods.

Methodology Overview

The framework combines three core components:

1. One-order Semantic Module: At the foundational level, semantic local features are extracted using a CNN backbone and a human key-points estimation model. This stage is primarily concerned with the raw semantic content captured in the images, but it involves adjustments to account for the innate challenges posed by occluded views.
2. High-order Relation Module: This innovative module uses the local features as nodes in a graph structure and applies a novel adaptive direction graph convolutional (ADGC) layer. The ADGC layer is designed to propagate meaningful features while suppressing misleading node connections due to noise or occlusion. This is achieved through an automatic, directed adjustment of the linkage’s direction and degree, ensuring that valuable relational information from neighboring nodes is utilized optimally.
3. High-order Topology Module: This module focuses on aligning features between different images by viewing the task as a graph matching problem. The proposed cross-graph embedded-alignment (CGEA) layer integrates high-order topology information into feature extraction and alignment. The distinct advantage of CGEA is its ability to utilize graph matching to account for node and edge correspondences, replacing one-to-one matches with a robust collective soft-matching strategy. The embedded alignment straightly predicts similarity scores across instances.

Empirical Evaluation

The efficacy of the proposed method is validated through comprehensive experimentation on occluded, partial, and holistic ReID datasets. Specifically, the evaluations conducted on Occluded-Duke demonstrate that the proposed method substantially outperforms existing approaches, achieving an increase in mAP scores of 6.5%. This improvement underscores the utility of incorporating high-order relational and topological information for occluded ReID scenarios.

Implications and Future Work

The paper contributes significantly to the ReID field by integrating relational and topological insights into occluded person re-identification frameworks, establishing a benchmark for future research. The use of an adaptive graph-based approach opens new opportunities for innovating in noise suppression and feature propagation techniques, which are critical under challenging view conditions. Moreover, the framework’s reliance on learned component interactions to facilitate nuanced feature alignment without strict one-to-one mapping paves the way for more error-tolerant systems in practical surveillance deployments.

Future directions encouraged by this research include exploring the intersection of graph convolutional networks with other domain-specific tasks requiring dynamic feature alignment, expanding the current model to tackle domain variances in different camera views and environments, and improving computational efficiency to ensure scalability to broader real-world applications. The continuous expansion of the framework could influence broader utilization of AI in various visual recognition settings beyond traditional surveillance, such as autonomous vehicles and human-robot interaction systems. The implications for AI-related privacy and ethical use also warrant thoughtful consideration as such technologies proliferate.