Beyond Pairwise Matching: Person Reidentification via High-Order Relevance Learning

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person reidentification (Re-ID) is challenging due to variations in illumination, viewpoint, background, and occlusion.
• Existing Re-ID methods often rely on pairwise distance metric learning, which can be suboptimal and limit data utilization.
• The correlation within gallery data offers valuable information for improving Re-ID performance.

Purpose of the Study:

• To develop a novel approach for person reidentification that overcomes limitations of pairwise matching.
• To investigate the utility of high-order correlations among probe and gallery data for learning relevance.
• To propose a flexible framework that integrates multiple features for enhanced person description.

Main Methods:

• A multihypergraph joint learning algorithm is proposed to model high-order correlations.
• Multiple hypergraphs are constructed, each utilizing a distinct feature type from the imaging data.
• The learning process operates on the multihypergraph structure, determining probe identity via relevance to gallery data.

Main Results:

• The proposed method effectively explores relationships among multiple images, moving beyond pairwise comparisons.
• Multimodal data, represented by different features, are integrated within the multihypergraph structure, enriching the learning process.
• Experimental results on three public datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The multihypergraph joint learning framework offers a flexible and powerful approach for person reidentification.
• This method enhances Re-ID by jointly learning from high-order correlations and multiple feature representations.
• The proposed scheme provides a general framework adaptable to various feature combinations, showing significant practical advantages.