srpone/look-bench

LookBench是一个大规模、开放的时尚图像检索基准数据集，旨在评估现代视觉和视觉语言模型在现实、污染感知环境下的表现。该数据集强调实时数据、领域多样性和涵盖单品及搭配级别的整体检索任务。它包含多个子集，如RealStudioFlat、AIGen-Studio、RealStreetLook和AIGen-StreetLook，每个子集具有不同的图像来源和检索难度。数据集支持单件物品检索和多件物品（搭配）检索两种主要任务，反映了现实世界的时尚搜索和推荐场景。每个子集包含查询分割和库分割，样本可能包括图像、时尚类别标签、边界框、产品ID、任务类型和难度等级等信息。

LookBench is a large-scale, open benchmark for fashion image retrieval, designed to evaluate modern vision and vision–language models under realistic, contamination-aware settings. The benchmark emphasizes live data, domain diversity, and holistic retrieval tasks spanning both single-item and outfit-level scenarios. It consists of multiple subsets such as RealStudioFlat, AIGen-Studio, RealStreetLook, and AIGen-StreetLook, each reflecting different image sources and retrieval difficulties. The dataset supports two primary retrieval tasks: single-item retrieval and multi-item (outfit) retrieval, mirroring real-world fashion search and recommendation scenarios. Each subset includes query and corpus splits, with samples potentially containing fields like image, category label, bounding box, item ID, task type, and difficulty level.