Cross-modal dual-branch parallel hybrid matching for text-to-image person retrieval

Text-to-image person retrieval aims to return pedestrian images that semantically match a natural-language query. We introduce Cross-modal Dual-branch Parallel Hybrid Matching (CDPHM), which couples a global CLIP branch with a lightweight local branch via bidirectional attention, fusing holistic semantics and fine-grained spatial cues in a shared embedding space. Extensive experiments on three benchmark datasets show that CDPHM achieves state-of-the-art Rank-5 performance on all three datasets, strong Rank-10 results, and the best overall performance on RSTPReid with 61.65% Rank-1. On CUHK-PEDES and ICFG-PEDES, CDPHM attains 73.89% and 63.92% Rank-1, respectively. Moreover, CDPHM maintains a compact model size of 191 M parameters, with only marginal parameter overhead compared with representative CLIP-based methods while effectively combining global and local matching. The code has already been made publicly available at https://github.com/mian-hu/CDPHM.