CLIP-Driven Fine-Grained Text-Image Person Re-Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Text-Image Person Re-identification (TIReID) aims to match images with text descriptions.
• Current methods struggle with multi-modal correspondence and fine-grained details.
• Vision-Language Pre-training models like CLIP offer potential but have limitations in capturing nuanced information for TIReID.

Purpose of the Study:

• To propose a novel CLIP-driven Fine-grained information excavation framework (CFine) for enhanced TIReID.
• To address the limitations of existing methods in leveraging multi-modal correspondence and fine-grained information.
• To improve the accuracy and efficiency of retrieving images based on text queries in TIReID.

Main Methods:

• Developed a CLIP-driven Fine-grained information excavation framework (CFine).
• Introduced a multi-level global feature learning (MGF) module to mine discriminative local information and enhance global features.
• Designed cross-grained feature refinement (CFR) and fine-grained correspondence discovery (FCD) modules for establishing cross-modal correspondence at multiple levels.

Main Results:

• The proposed CFine framework effectively utilizes CLIP's knowledge for TIReID.
• The MGF module enhances the extraction of identity-related discriminative clues.
• Experiments on multiple benchmarks demonstrate superior performance compared to existing methods.

Conclusions:

• CFine significantly improves Text-Image Person Re-identification performance.
• The framework successfully addresses the challenges of fine-grained information and cross-modal correspondence.
• CFine offers a promising direction for advancing TIReID research.