Combination of Multiple Alignment Losses and Mutual Noise Filtering for Text-Based Person Search

Text-based person search (TPS) aims to retrieve target individuals based on textual queries. Despite the considerable research in this area, a fundamental challenge remains unresolved: learning to map visual and textual modalities into a shared feature space. Additionally, noisy datasets and variations in language within caption queries pose significant obstacles. To address these issues, we introduce CALM-TPS, a novel approach for TPS that combines multiple alignment losses and data filtering. Our contributions include: (1) optimizing representations by different alignment losses and (2) employing mutual noise filtering to identify valuable samples within datasets. We also validate our method's effectiveness through rigorous experiments on three popular benchmarks: CUHK-PEDES, ICFG-PEDES, and RSTPSeid, demonstrating superior performance over recent methods.