Browse Teach Curate

Part-Level Cross-Modal Correspondence for Text-Based Person Search

How to match free-form text descriptions to person images by aligning fine-grained body-part information across the image and text modalities. Students learn how a coarse-to-fine attention-based encoder-decoder aligns the two modalities without explicit alignment labels, and how a commonality-based margin ranking loss learns discriminative part details using only person-ID supervision.

Data Text

2501.00318

This paper addresses text-based person search: retrieving the person images most relevant to a free-form natural-language description, a task important for surveillance and public-safety applications…