Title: ROGLE: Robust Global-Local Alignment with Automated Region Supervision for Text-Based Person Search

Abstract:

Text-Based Person Search (TBPS) is designed to locate pedestrian images by utilizing natural language descriptions. Despite its utility, current TBPS systems—particularly those leveraging CLIP architectures—face significant hurdles in achieving fine-grained comprehension. This limitation stems from global representational bias and semantic sparsity, issues inherent to models trained on brief captions. Consequently, these systems exhibit poor fine-grained alignment, a problem compounded by the limited availability of region-level annotations.

To resolve these challenges, we introduce ROGLE (Robust Global-Local Embedding), a comprehensive framework that eliminates the need for expensive manual labeling via an Automated Region-to-Sentence Matching (RSM) strategy. RSM efficiently generates pseudo region-sentence pairs, enabling scalable fine-grained supervision. Additionally, ROGLE utilizes a multi-granular learning approach that integrates global contrastive learning with local alignment at the region level.

We also present the P-VLG Benchmark, a new large-scale dataset created by curating and enhancing images from existing public benchmarks. Featuring more than 100,000 annotated regions and detailed long-form captions, P-VLG is the first TBPS benchmark capable of supporting both global and local evaluation protocols. Our extensive experiments demonstrate that ROGLE substantially surpasses current methods, with notable improvements on difficult long-form queries. Both the source code and the P-VLG benchmark dataset will be released to the public.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC