M2ReID: Multimodal and Multi-task Person Re-identification Dataset

Person re-identification (ReID) serves as a cornerstone in applications such as security surveillance and criminal investigations. However, most existing datasets are constrained to cropped pedestrian images and unimodal queries, which overlook the complexity of real-world scenes involving multiple individuals, occlusions, and background clutter. Moreover, multimodal resources that jointly support image- and text-based retrieval, as well as pixel-level segmentation, remain scarce. To address these limitations, we present M\u00b2ReID, currently the largest full-scene multimodal dataset for person ReID and segmentation. M\u00b2ReID contains over 200K images covering 4,894 identities, each paired with high-quality segmentation masks and fine-grained textual descriptions. It is constructed through a hybrid annotation pipeline, where manually annotated bounding boxes are used as spatial prompts to guide SAM-2 for pixel-level mask generation, while GPT-4o produces descriptive captions under a carefully designed attribute taxonomy, followed by human refinement. This design ensures both scalability and annotation fidelity. By providing multimodal queries and unified benchmarks for retrieval and segmentation, M\u00b2ReID fills a critical gap in existing resources and establishes a solid foundation for advancing multimodal human-centric understanding in complex real-world scenarios.