Improving Text-Based Person Retrieval by Excavating All-Round Information Beyond Color

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Text-based person retrieval relies on searching visual data using text queries.
• Current methods often overemphasize color (CLR), neglecting crucial texture and structure cues, leading to suboptimal performance.
• This over-reliance on color information limits the effectiveness of person retrieval systems.

Purpose of the Study:

• To propose a novel framework, EAIBC (Excavate All-round Information Beyond Color), to improve text-based person retrieval.
• To address the limitations of color over-reliance in existing person retrieval approaches.
• To enhance the extraction and utilization of diverse visual features beyond color.

Main Methods:

• The EAIBC framework incorporates four distinct branches: RGB, grayscale (GRS), high-frequency (HFQ), and CLR.
• A mutual learning (ML) mechanism is introduced to enable effective communication and balanced information integration among the branches.
• The framework is designed to leverage all-round visual information comprehensively.

Main Results:

• EAIBC was evaluated on three benchmark datasets: CUHK-PEDES, ICFG-PEDES, and RSTPReid.
• The proposed method demonstrated significant performance improvements over existing approaches.
• EAIBC achieved state-of-the-art (SOTA) results across supervised, weakly supervised, and cross-domain retrieval settings.

Conclusions:

• The EAIBC framework effectively mitigates the problem of color over-reliance in text-based person retrieval.
• By integrating multi-modal visual information through mutual learning, EAIBC achieves superior retrieval accuracy.
• The proposed method represents a significant advancement in text-based person retrieval, offering robust performance across various scenarios.