SpaceVista/SpaceVista-Bench

SpaceVista是一个用于全尺度视觉空间推理的评估基准数据集，覆盖从毫米到公里的空间范围。该数据集包含四个主要类别：TinyTabletop（毫米尺度，细粒度空间细节的桌面小物体）、Tabletop（厘米尺度，标准桌面场景）、Indoor（米尺度，室内房间环境）和Outdoor（公里尺度以下，大规模户外和驾驶场景）。数据以JSON格式（bench_final.json）统一组织，包含所有评估条目，并附带提取的视频帧（PNG格式）和原始视频（MP4格式）。该数据集支持多模态任务，如视频文本到文本，适用于空间推理、机器人、自动驾驶和具身智能体等领域的研究和评估。SpaceVista is an evaluation benchmark for all-scale visual spatial reasoning from millimeters to kilometers. The dataset includes four main categories: TinyTabletop (mm scale, small-scale tabletop objects with fine-grained spatial details), Tabletop (cm scale, standard tabletop scenes), Indoor (m scale, room-level indoor environments), and Outdoor (<km scale, large-scale outdoor and driving scenes). The data is organized in a unified JSON format (bench_final.json) containing all evaluation entries, along with extracted video frames (PNG) and raw videos (MP4). It supports multimodal tasks such as video-text-to-text and is designed for applications in spatial reasoning, robotics, autonomous driving, and embodied agents.