Text-based person retrieval (TBPR) aims to search for target person images from large-scale video clips or image databases based on textual descriptions. The quality of benchmarks is critical to accurately evaluating TBPR models for their ability in relation to cross-modal matching. However, we find that existing TBPR benchmarks have a common problem, which often leads to ambiguities where multiple images of persons with different identities have very similar or even identical textual descriptions. As a consequence, although TBPR models correctly retrieve the images corresponding to a given description, such matches may be erroneously evaluated as mismatches due to the above annotation problem. We argue that the main cause of this problem is that each person image is annotated individually without reference to other similar images, making it challenging to provide distinctive descriptions for each image. To address this problem, we propose an effective and efficient annotation refinement framework to improve the annotation quality of TBPR benchmarks and thereby mitigate annotation-induced mismatches. Firstly, sets of images prone to mismatches are automatically identified by TBPR models. Then, by leveraging multimodal large language models (MLLMs), multiple images are simultaneously processed and distinctive descriptions are generated for each image. Finally, the original descriptions are replaced to improve the annotation quality. Extensive experiments on three popular TBPR benchmarks (CUHK-PEDES, RSTPReid and ICFG-PEDES) validate the effectiveness of our proposed method for improving the quality of annotations, and demonstrate that the resulting more discriminative captions can truly benefit the mainstream TBPR models. The improved annotations of these benchmarks will be released publicly.

The task of pedestrian attribute recognition (PAR) identifies a set of predefined attributes in pedestrian images from surveillance videos or collected imagery, which are often adopted as important mid-level features in higher-level tasks, such as person re-identification, pedestrian detection, etc. In these cases, the domain differences between datasets of different tasks will lead to clear performance degradation of the mainstream PAR methods. This degradation becomes significant in the application of remote sensor-based PAR, since the model is trained on traditional fixed-camera visual data while applied on UAV-based remote sensor data, facing more cross-domain challenges. To address these issues, we formally introduce in this paper the task of cross-domain pedestrian attribute recognition (CD_PAR) for the first time, and efficiently establish a set of evaluation criteria for this new task. In addition, to facilitate the future research of CD_PAR, we propose a new baseline method named local domain discriminator-based cross-domain pedestrian attribute recognition (LDCD_PAR), by introducing a local domain discriminator based on adversarial training to effectively obtain the fine-grained domain-invariant features. Extensive well designed cross-domain experimental evaluation and application on remote sensor-based PAR demonstrate the value of the new CD_PAR task, and validate the effectiveness of our new baseline method.