Text-to-image person re-identification (TIReID) aims to retrieve the corresponding pedestrian image based on a text description. Prior research primarily focused on intra-domain TIReID task, where training and testing data are drawn from the same domain. However, they disregard domain gaps among various TIReID datasets, leading to performance degradation in cross-domain scenarios. To investigate cross-domain problem, this paper analyzes the causes of differences across datasets and empirically proves that the text descriptions from different annotators exhibit significant style discrepancies, even when annotating the same images. Motivated by this insight, we propose a benchmark with Unified Text-Domain (UniTD) to reduce the domain gap. Specifically, we harness the power of MLLMs to generate textual descriptions with a unified style. Moreover, considering the domain differences between UniTD text and real-world query text, we present a Query Text Rewriting method (QTR). It utilizes LLMs to rewrite the query text by chain-of-thought, aiming to transfer the style of query text into UniTD style. Extensive experiments demonstrate that the proposed UniTD benchmark, with its unified text-domain, significantly improves cross-domain performance by 11%-33%. Additionally, QTR effectively aligns the style of real-world query text with UniTD, achieving state-of-the-art cross-domain performance. The code and datasets are available in https://github.com/laiping-lp/UniTD.git.

Adapting large-scale image-text pre-training models, e.g., CLIP, to the video domain represents the current state-of-the-art for text-video retrieval. The primary approaches involve transferring text-video pairs to a common embedding space and leveraging cross-modal interactions on specific entities for semantic alignment. Though effective, these paradigms entail prohibitive computational costs, leading to inefficient retrieval. To address this, we propose a simple yet effective method, Global-Local Semantic Consistent Learning (GLSCL), which capitalizes on latent shared semantics across modalities for text-video retrieval. Specifically, we introduce a parameter-free global interaction module to explore coarse-grained alignment. Then, we devise a shared local interaction module that employs several learnable queries to capture latent semantic concepts for learning fine-grained alignment. Furthermore, an Inter-Consistency Loss (ICL) is devised to accomplish the concept alignment between the visual query and corresponding textual query, and an Intra-Diversity Loss (IDL) is developed to repulse the distribution within visual (textual) queries to generate more discriminative concepts. Extensive experiments on five widely used benchmarks (i.e., MSR-VTT, MSVD, DiDeMo, LSMDC, and ActivityNet) substantiate the superior effectiveness and efficiency of the proposed method. Remarkably, our method achieves comparable performance with SOTA as well as being nearly 220 times faster in terms of computational cost. Code is available at: https://github.com/zchoi/GLSCL.