InternScenes
收藏资源简介:
InternScenes是一个包含约40,000个室内场景的大规模可模拟数据集,这些场景具有多样化的布局和真实感。该数据集整合了来自三个不同来源的场景数据:真实世界的扫描场景、程序生成的场景和设计师创作的场景。数据集包含1.96M个3D对象,涵盖了15种常见的场景类型和288个对象类别。InternScenes的特点是场景多样性和真实布局,每个区域平均包含41.5个对象,包括大量小物品,从而形成了复杂且真实的布局。该数据集通过创建真实场景的模拟副本、增强交互性、并通过物理模拟解决对象碰撞问题,确保了可模拟性。InternScenes已被用于两个基准应用:场景布局生成和点目标视觉导航,为具身AI的训练和模型训练提供了重要资源。
InternScenes is a large-scale physically simulable dataset containing approximately 40,000 indoor scenes with diverse layouts and high realism. This dataset integrates scene data from three distinct sources: real-world scanned scenes, procedurally generated scenes, and designer-created scenes. The dataset comprises 1.96 million 3D objects, spanning 15 common scene types and 288 object categories. Distinguished by its diverse scenes and realistic layouts, each scene on average contains 41.5 objects, with a substantial number of small items included, thereby forming complex and authentic layouts. This dataset ensures its simulability by generating simulated replicas of real-world scenes, enhancing interactivity, and resolving object collision issues via physical simulation. InternScenes has been utilized in two benchmark applications: scene layout generation and point-goal visual navigation, serving as a critical resource for embodied AI training and model development.




