Reasoning and Tuning: Graph Attention Network for Occluded Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification (re-id) is challenging due to missing visual data.
• Existing methods often discard occluded parts, leading to significant semantic loss and reduced matching confidence.

Purpose of the Study:

• To develop a novel approach for occluded person re-identification that compensates for semantic loss.
• To improve the accuracy of matching partially occluded individuals by leveraging holistic image information.

Main Methods:

• Proposed the Reasoning and Tuning Graph Attention Network (RTGAT) for learning complete person representations.
• Self-mined semantic correlations to reason body part visibility scores.
• Utilized visibility scores as graph attention within a Graph Convolutional Network (GCN) to suppress noise and propagate missing information.

Main Results:

• RTGAT effectively learns complete person representations from occluded images.
• The method demonstrated superior performance on occluded person re-identification benchmarks.
• Joint reasoning of visibility and compensation significantly improved feature matching confidence.

Conclusions:

• The proposed RTGAT method offers a robust solution for occluded person re-identification.
• Compensating for semantic loss using holistic image counterparts is a promising strategy.
• The approach enhances the ability to match occluded individuals accurately.