purespace
收藏资源简介:
PureSpace是一个用于评估视觉语言模型(VLMs)抽象空间推理能力的基准数据集,专注于几何推理任务,要求模型理解三维物体的空间变换,例如旋转、投影视图(如俯视图)以及立方体补全。数据集中包含三种核心任务:旋转(rotation)、投影(projection)和补全(completion)。每个任务样本由一个提问图像、一组困难选项图像(及其对应的正确答案索引)、一组简单选项图像(及其对应的正确答案索引)以及相关的元数据文件构成。提问文本根据任务类型而定,例如“哪个选项是给定物体的旋转版本?”。数据集以结构化的目录形式组织,包含images/和labels/两个主要目录,labels/下按任务和训练/测试划分存储文本标签文件。该数据集旨在为视觉语言模型的空间智能提供一个纯净、具有挑战性的评估基准。
PureSpace is a benchmark dataset for evaluating the abstract spatial reasoning capabilities of vision-language models (VLMs). It focuses on geometric reasoning tasks, requiring models to understand spatial transformations of 3D objects, such as rotation, projection views (e.g., top views), and cube completion. The dataset includes three core tasks: rotation, projection, and completion. Each task sample consists of a query image, a set of hard option images (with corresponding correct answer indices), a set of easy option images (with corresponding correct answer indices), and related metadata files. The query text varies based on the task type, e.g., "Which option is the rotated version of the given object?". The dataset is organized in a structured directory format, with two main directories: images/ and labels/, where labels/ stores text label files divided by task and train/test splits. It aims to provide a pure and challenging evaluation benchmark for the spatial intelligence of vision-language models.
- 数据集名称: PureSpace
- 许可协议: cc-by-nc-4.0
- 任务类型: 视觉问答 (visual-question-answering)、图像到文本 (image-text-to-text)
- 语言: 英文 (en)
- 标签: 空间智能 (spatial-intelligence)、几何推理 (geometric-reasoning)、基准测试 (benchmark)
- 简介: PureSpace 是一个用于评估视觉-语言模型在抽象空间推理能力上的基准测试数据集。
- 数据集结构:
images/: 包含多层级的图像文件,例如l3_c221/000009/000009_iso.jpg、000009_top.jpg等。labels/: 包含三个子任务目录:rotation/: 旋转任务projection/: 投影任务completion/: 补全任务- 每个子任务下再分为
train/和test/文件夹,内部为.txt标签文件。
- 使用示例: 通过读取标签文件,可以获取每个样本的问题图像、硬选项/易选项图像及其对应的答案索引,以及元数据路径。
- 引用: 论文题为 "PureSpace: A Benchmark for Abstract Spatial Reasoning in Vision-Language Models",收录于 CVPR 2026 Findings,作者为 Li, Jinkai 等。




