Joint Person Objectness and Repulsion for Person Search

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person search in unconstrained images combines detection and matching.
• Existing methods overlook objectness and repulsion, crucial for handling distractors.

Purpose of the Study:

• To propose a novel OR similarity metric incorporating objectness and repulsion for improved person search.
• To reduce the impact of distractor images in person search tasks.

Main Methods:

• Developed an OR similarity metric combining visual similarity, objectness, and repulsion terms.
• Integrated OR similarity into a Detection-Matching framework using Faster R-CNN.
• Evaluated on PRW and CUHK-SYSU datasets with six description models.

Main Results:

• The OR similarity effectively reduces the influence of distractor samples.
• Significant performance boosts observed: mAP increased from 92.32% to 93.23% on CUHK-SYSU.
• mAP improved from 50.91% to 52.30% on PRW dataset.

Conclusions:

• The proposed OR similarity metric enhances person search by effectively filtering distractors.
• Jointly considering objectness and repulsion leads to superior performance in person search systems.