SpaCE-10
收藏资源简介:
SpaCE-10是一个用于评估多模态大语言模型在室内环境中的组合空间智能的基准数据集。它涵盖了10种原子空间能力、8种组合QA类型、5000多个QA对和811个室内场景(ScanNet++、ScanNet、3RScan、ARKitScene),同时支持2D和3D MLLMs的评估。
SpaCE-10 is a benchmark dataset designed to evaluate the combined spatial intelligence of multimodal large language models within indoor environments. It encompasses 10 atomic spatial abilities, 8 types of composite question-answering, over 5,000 QA pairs, and 811 indoor scenes (ScanNet++, ScanNet, 3RScan, ARKitScene), and supports the evaluation of both 2D and 3D MLLMs.
SpaCE-10 数据集概述
基本信息
- 数据集名称: SpaCE-10
- 用途: 评估多模态大语言模型(MLLMs)在室内环境中的组合空间智能
- 主要特点:
- 包含10种原子空间能力
- 8种组合问答类型
- 5,000+问答对
- 811个室内场景(来自ScanNet++、ScanNet、3RScan、ARKitScene)
- 支持2D和3D MLLMs评估
数据集内容
- 场景来源: ScanNet++, ScanNet, 3RScan, ARKitScene
- 数据类型: 问答对
- 数量: 5,000+ QA pairs
评估方法
- 评估工具: lmms-eval
- 环境要求: Python 3.10
- 示例评估命令: bash bash internvl2.5-8b.sh
最新动态
- 2025/06/09: 3D MLLMs的扫描数据和手动收集的3D快照即将发布
- 2025/06/09: 评估代码已发布
- 2025/06/08: 2D MLLMs的基准测试已发布
引用信息
bibtex @article{gong2025space10, title={SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence}, author={Ziyang Gong, Wenhao Li, Oliver Ma, Songyuan Li, Jiayi Ji, Xue Yang, Gen Luo, Junchi Yan, Rongrong Ji}, journal={arXiv preprint arXiv:XXXX.XXXXX}, year={2025} }
相关链接
- 项目页面: https://github.com/VisionXLab/SpaCE-10
- HuggingFace数据集: https://huggingface.co/datasets/Cusyoung/SpaCE-10




