Text-Based Person Search aims to retrieve target pedestrian images from large galleries using natural language descriptions. This task is of great importance in public security and surveillance. However, it remains highly challenging due to modality discrepancy, fine-grained inter-identity variations, and insufficient semantic alignment. To tackle these challenges, we propose a unified Global-Local Alignment Framework that jointly strengthens global distributional alignment and local discriminability. In the global branch, we introduce three complementary components: an improved similarity distribution matching module that refines alignment with soft labels and temperature scaling, a stochastic pairwise matching module that balances strong and weak positives to mitigate noisy supervision, and a hard negative contrastive matching module that emphasizes boundary discrimination by mining the most confusing negatives. In the local branch, we design two lightweight yet effective modules: a local identity discrimination module, which enforces identity supervision at the slot level, and a discriminative local matching module, which employs dynamic slot weighting for contrastive learning. Extensive experiments on three public benchmarks: CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that our method outperforms prior approaches in retrieval accuracy.
Paper
Full text
GLAF: Global and Local Alignment Framework for Text-Based Person Search
Semantic Scholar · 2025
Abstract
Text-Based Person Search aims to retrieve target pedestrian images from large galleries using natural language descriptions. This task is of great importance in public security and surveillance. However, it remains highly challenging due to modality discrepancy, fine-grained inter-identity variations, and insufficient semantic alignment. To tackle these challenges, we propose a unified Global-Local Alignment Framework that jointly strengthens global distributional alignment and local discriminability. In the global branch, we introduce three complementary components: an improved similarity distribution matching module that refines alignment with soft labels and temperature scaling, a stochastic pairwise matching module that balances strong and weak positives to mitigate noisy supervision, and a hard negative contrastive matching module that emphasizes boundary discrimination by mining the most confusing negatives. In the local branch, we design two lightweight yet effective modules: a local identity discrimination module, which enforces identity supervision at the slot level, and a discriminative local matching module, which employs dynamic slot weighting for contrastive learning. Extensive experiments on three public benchmarks: CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that our method outperforms prior approaches in retrieval accuracy.