Habitat-Matterport 3D Semantics (HM3DSEM)
收藏资源简介:
HM3DSEM数据集是由Meta AI等机构创建的,是目前学术界可用的最大3D真实世界空间密集语义标注数据集。该数据集包含216个高分辨率3D扫描场景,覆盖3100个房间,总计142,646个对象实例标注。数据集的创建过程涉及大量人工标注和验证,总计超过14,200小时。HM3DSEM数据集主要用于提升下游具身AI任务的性能,特别是在Object Goal Navigation任务中,通过不同学习方法训练的策略在跨数据集泛化性能上均有显著提升。
The HM3DSEM dataset, developed by Meta AI and other institutions, stands as the largest densely annotated 3D semantic dataset of real-world spaces currently available in academia. This dataset comprises 216 high-resolution 3D scanned scenes spanning 3,100 rooms, with a total of 142,646 annotated object instances. The development of this dataset involved extensive manual annotation and verification, totaling over 14,200 hours. The HM3DSEM dataset is primarily used to enhance the performance of downstream embodied AI tasks. Notably, in the Object Goal Navigation task, policies trained using various learning approaches have all demonstrated substantial improvements in cross-dataset generalization capabilities.



