[Paper Review] High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification

Guan’an Wang, Shuo Yang+7 more|arXiv (Cornell University)|Mar 18, 2020

Video Surveillance and Tracking Methods42 references43 citations

TL;DR

Introduces a high-order information framework for occluded ReID by modeling intra-image relations with adaptive directed graph convolutions and cross-image topology via graph matching to produce robust similarity estimation.

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.

Motivation & Objective

• Motivate occluded person ReID as a problem with significant occlusion and misalignment challenges.
• Propose a framework that learns high-order relation among local features via a graph-based approach.
• Propose a cross-graph topology module to robustly align two images without hard one-to-one matching.
• Demonstrate that high-order information yields significant performance gains on occluded, partial, and holistic ReID datasets.

Proposed method

• Extract semantic local features from key-point regions using a CNN backbone and a key-points estimation model.
• Model local features as graph nodes and apply an adaptive-directed graph convolutional (ADGC) layer to learn and suppress meaningless message passing.
• Align two images by a cross-graph embedded-alignment (CGEA) layer that embeds topology information via graph matching and predicts similarity.
• End-to-end training with a verification loss to supervise similarity prediction between image pairs.
• Fuse relation-based similarity and topology-based similarity into a final scalar score for ranking.

Experimental results

Research questions

• RQ1 Can high-order relation information improve discrimination in occluded ReID by repairing occluded region features through graph-based message passing?
• RQ2 Does cross-graph topology learning via graph matching provide robust, soft alignment that outperforms hard one-to-one matching in occluded/partial scenarios?
• RQ3 Do the proposed ADGC and CGEA layers provide measurable gains on occluded, partial, and holistic ReID benchmarks?
• RQ4 How does the full HOReID framework perform across occluded, partial, and holistic datasets compared to state-of-the-art methods?

Key findings

• HOReID outperforms state-of-the-art on Occluded-Duke by delivering a Rank-1 of 55.1 and mAP of 43.8.
• On Occluded-ReID, HOReID achieves a Rank-1 of 80.3 and mAP of 70.2.
• Ablation shows benefits from adding semantic features, high-order relation, and topology modules (Rank-1: 49.9 → 55.1).
• ADGC effectively suppresses noisy features while promoting meaningful ones, improving robustness.
• CGEA enables robust, soft cross-image alignment and improves final similarity predictions.