LAttE: A label-free and multimodal framework for context-aware person re-identification

Dasom Ahn, Sangwon Kim, Kwang-Ju Kim, Byoung Chul Ko

Abstract: Highlights•Label-free Re-ID using auto attributes and CLIP for scalable, explainable matching.•Attribute bank via GPT-4o and CLIP expands beyond predefined label space.•CLIP-based similarity matrix provides pseudo-labels for attribute supervision.•Cross-attention fusion with bone tokens for joint visual-pose attribute learning.