P-CLIP: Progressive Discrepancy Learning for One-Shot Text-to-Image Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Supervised text-to-image person re-identification (TIReID) requires extensive annotated data, limiting its practicality for large-scale camera networks.
• One-shot TIReID offers a solution by reducing annotation requirements, using only a single labeled image-text pair per identity.
• A key challenge is establishing robust visual-textual correspondences across diverse viewing conditions without paired cross-view data.

Purpose of the Study:

• To develop a progressive discrepancy learning framework (P-CLIP) for one-shot TIReID.
• To create a unified embedding space resilient to view-specific biases.
• To address the challenge of learning cross-view visual-textual correspondences with minimal supervision.

Main Methods:

• Proposed a Progressive Multi-View Generation (MVG) method to create multiple noisy views from a single labeled instance.
• Introduced a Cross-View Discrepancy Learning (CDL) module to mitigate ambiguities by leveraging inter-view discrepancies.
• Developed a Compact Cross-Modal Matching (CCM) loss to enhance correspondence learning by emphasizing matched pairs and suppressing unmatched ones.

Main Results:

• The P-CLIP framework demonstrated significant effectiveness in one-shot TIReID tasks.
• Experimental results on three benchmark datasets validated the proposed method's performance.
• The approach successfully integrated multimodal error correction into person re-identification.

Conclusions:

• The proposed P-CLIP framework effectively addresses the challenges of one-shot TIReID.
• The method significantly reduces the annotation burden for practical large-scale person re-identification systems.
• The study provides a robust approach for learning cross-view visual-textual correspondences in person re-identification.