HM3D

Habitat-Matterport 3D Dataset（HM3D）是由Facebook AI Research联合Matterport公司创建的大规模室内三维环境数据集，旨在为具身人工智能研究提供高质量的三维场景数据。该数据集包含1000个真实世界的建筑级三维重建场景，涵盖多层住宅、商店及其他室内空间，总可导航面积达112,500平方米，是现有数据集中规模最大、重建完整性和视觉保真度最高的。数据集的创建过程包括从38个国家的181个地理区域采集三维扫描数据，通过Matterport Pro2深度传感器获取RGB-D数据，并利用Matterport重建流程生成纹理化三维网格。经过两阶段的注释和验证，最终筛选出1000个高质量场景。HM3D数据集在具身AI领域具有重要应用价值，可用于训练和评估视觉导航、语义理解等任务，助力AI代理在复杂室内环境中实现高效导航与交互。

Habitat-Matterport 3D Dataset (HM3D) is a large-scale indoor 3D environment dataset co-developed by Facebook AI Research and Matterport, aimed at providing high-quality 3D scene data for embodied artificial intelligence (embodied AI) research. This dataset contains 1,000 real-world building-level 3D reconstructed scenes, covering multi-story residences, commercial shops and other indoor spaces, with a total navigable area of 112,500 square meters. It is the largest existing dataset that features the highest reconstruction completeness and visual fidelity. The creation of HM3D involved collecting 3D scanning data from 181 geographic regions across 38 countries, acquiring RGB-D data via Matterport Pro2 depth sensors, and generating textured 3D meshes through Matterport's reconstruction pipeline. Following a two-stage annotation and validation process, 1,000 high-quality scenes were finally selected. The HM3D dataset holds significant application value in the embodied AI field, and can be used to train and evaluate tasks such as visual navigation and semantic understanding, enabling AI Agents to achieve efficient navigation and interaction in complex indoor environments.

搜集汇总

背景与挑战

背景概述

HM3D是当前最大的3D室内空间数据集，包含1000个从真实环境生成的高分辨率3D扫描，覆盖住宅、商业和公共空间，适用于学术研究。它免费提供，可与Habitat模拟器结合训练具身智能体（如家庭机器人和AI助手），并附带详细的元数据以支持场景分析和评估。