Relevance Metric Learning for Person Re-Identification by Exploiting Listwise Similarities

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person re-identification is crucial for video surveillance but challenging.
• Existing metric learning methods often use sparse pairwise constraints, discarding valuable information.
• This can lead to suboptimal performance in matching individuals across camera views.

Purpose of the Study:

• To propose a novel relevance metric learning method with listwise constraints (RMLLCs) for improved person re-identification.
• To capture comprehensive pairwise similarities using listwise similarity lists.
• To address limitations in capturing relative relevance by introducing a rectification term.

Main Methods:

• Developed a relevance metric learning method with listwise constraints (RMLLCs).
• Utilized listwise similarities to capture all pairwise similarities.
• Introduced a rectification term and an alternating iterative algorithm to learn the metric and exploit relative similarities.

Main Results:

• The proposed RMLLC method significantly outperforms state-of-the-art approaches on four benchmark datasets.
• The inclusion of the rectification term further enhances the performance of the metric learning method.
• Demonstrated superior accuracy in person re-identification tasks.

Conclusions:

• The novel listwise constraint-based metric learning method offers a more discriminative approach to person re-identification.
• Automatically exploiting relative similarities via a rectification term leads to substantial performance gains.
• The proposed method represents a significant advancement in robust person matching for surveillance.