idekoh/BeyondSingleObject
收藏资源简介:
BeyondSingleObject数据集是一个用于学习3D关系的大型语言模型数据集,包含多个基准测试。数据集涵盖以下内容:MO3D(多对象位置、比较和整体问答)、Shape Mating(几何对选择和推理)、Change Captioning(验证和差异描述)以及ModelNet40(零样本分类评估)。数据集中包含注释文件和PointLLM兼容的ModelNet40测试文件,但不包括点云资产;注释引用点云的相对路径,用户需从现有设置中重用点云根目录以解析路径。该数据集用于评估和训练,支持多对象3D关系理解任务。
The BeyondSingleObject dataset is a large language model dataset for learning 3D relations, comprising multiple benchmarks. It covers: MO3D (multi-object positional, comparative, and holistic QA), Shape Mating (geometric pair selection and reasoning), Change Captioning (verification and delta-captioning), and ModelNet40 (zero-shot classification evaluation). The dataset includes annotation files and a PointLLM-compatible ModelNet40 test file but does not include point-cloud assets; annotations reference point clouds by relative paths, requiring users to reuse point-cloud roots from existing setups to resolve paths. It is designed for evaluation and training, supporting multi-object 3D relation understanding tasks.




