rbler/MMSI-Video-Bench

MMSI-Video-Bench是一个全面的视频空间智能基准测试数据集，包含与每个问题对应的关键帧、原始视频数据（可选）、问题描述中使用的参考图像以及MMSI-Video-Bench注释。每个样本包含问题ID、问题类型（如空间构建、运动理解、规划、预测、跨视图推理）、参考图像列表、原始问题文本（可能包含<image>等占位符）、选项列表、关键帧列表、视频段信息列表、系统提示、任务提示、用户提示、格式提示和正确答案。数据集还提供了评估指南和不同模型在各种设置和子基准测试中的排行榜结果。

MMSI-Video-Bench is a holistic benchmark for video-based spatial intelligence, containing key frames corresponding to each question, original video data (optional), reference images used in the question descriptions, and MMSI-Video-Bench annotations. Each sample includes the question ID, type of the question (e.g., Spatial Construction, Motion Understanding, Planning, Prediction, Cross-View Reasoning), list of reference images, original question text (which may contain placeholders like <image>), list of options, list of key frames, list of video segment information, system prompt, task prompt, user prompt, format prompt, and ground truth. The dataset also provides evaluation guidelines and leaderboard results for different models across various settings and sub-benchmarks.