cvis-tmu/spar-vero-rl

SPAR-Vero-RL是一个用于空间推理任务的数据集，由SPAR-7M和Vero空间数据集组合而成。数据集包含训练集118,000个样本和测试集1,000个样本，涵盖多种任务类型，如外观顺序、相机运动推断、深度预测、距离推断、对象计数、空间关系、位置匹配、房间大小、空间想象、空间自监督学习任务（如相机空间、深度排序、翻转检测、拼图等）、空间体积推断、视图变化推断和3D视觉拼图等。这些任务旨在支持空间理解和推理能力的研究与应用。

SPAR-Vero-RL is a dataset for spatial reasoning tasks, combined from the SPAR-7M and Vero spatial datasets. It includes 118,000 training samples and 1,000 test samples, covering various task types such as appearance order, camera motion inference, depth prediction, distance inference, object counting, spatial relations, position matching, room size, spatial imagination, spatial self-supervised learning tasks (e.g., camera spatial, depth ordering, flip detection, jigsaw puzzles), spatial volume inference, view change inference, and 3D visual jigsaw. These tasks are designed to support research and applications in spatial understanding and reasoning.