InHabitants
收藏资源简介:
InHabitants是由蒂宾根大学人工智能中心等机构联合创建的大规模3D人-场景交互数据集,包含约78,000个样本,覆盖800个建筑级场景。该数据集通过创新的渲染-生成-提升流程自动生成,整合了3D场景几何、SMPL-X人体模型和RGB图像,实现了场景感知的多样化人体交互合成。数据来源于Habitat-Matterport3D数据集,经过视觉语言模型和图像编辑模型的联合处理,最终优化为物理合理的3D人体姿态。该数据集旨在解决3D场景理解中真实人-环境交互数据稀缺的问题,可广泛应用于机器人导航、虚拟现实和计算机视觉等领域。
InHabitants is a large-scale 3D human-scene interaction dataset jointly created by the AI Center of the University of Tübingen and other institutions. It contains approximately 78,000 samples covering 800 building-level scenes. The dataset is automatically generated through an innovative render-generate-enhance pipeline, which integrates 3D scene geometry, SMPL-X human body models and RGB images to realize scene-aware diverse human-scene interaction synthesis. Derived from the Habitat-Matterport3D dataset, the dataset is jointly processed by vision-language models and image editing models, and finally optimized into physically plausible 3D human poses. This dataset aims to address the scarcity of real human-environment interaction data in 3D scene understanding, and can be widely applied to fields such as robot navigation, virtual reality and computer vision.
InHabit数据集概述
数据集名称
InHabit: Leveraging Image Foundation Models for Scalable 3D Human Placement
作者与机构
- 作者:Nikita Kister, Pradyumna YM, Istvan Sarandi, Jiayi Wang, Anna Khoreva, Gerard Pons-Moll
- 机构:University of Tübingen, Tübingen AI Center, Bosch Center of AI, Zalando SE, Max Planck Institute for Informatics, Saarland Informatics Campus
核心方法
- 遵循渲染-生成-提升原则,实现3D场景中交互式人类的自动、可扩展生成。
- 给定渲染的3D场景,通过视觉语言模型提出上下文有意义的动作,图像编辑模型插入人类,优化过程将编辑结果提升为与场景几何对齐的物理合理SMPL-X人体。
数据集内容
- 规模:78K样本,覆盖约800个建筑级场景。
- 数据类型:包含完整的3D几何、SMPL-X人体模型和RGB图像。
- 应用场景:用于训练具身智能体理解3D场景中的人类交互。
应用效果
- 增强标准训练数据后,可改进基于RGB的3D人-场景重建和接触估计。
- 在感知用户研究中,78%的情况下优于现有技术。
相关资源
-
论文/预印本:提供链接(原文中标记为"Paper __ Preprint")。
-
代码与数据:提供链接(原文中标记为"Code Data")。
-
BibTeX引用:
@misc{inhabit2026, author = {Kister, Nikita and YM, Pradyumna and S{a}r{a}ndi, Istv{a}n and Wang, Jiayi and Khoreva, Anna and Pons-Moll, Gerard}, title = {InHabit: Leveraging Image Foundation Models for Scalable 3D Human Placement}, year = {2026}, howpublished = {url{https://virtualhumans.mpi-inf.mpg.de/inhabit/}}, }
数据集地址
https://virtualhumans.mpi-inf.mpg.de/inhabit/

- 1InHabit: Leveraging Image Foundation Models for Scalable 3D Human Placement蒂宾根大学·人工智能中心; 博世人工智能中心; Zalando SE; 马克斯·普朗克信息学研究所 · 2026年



