SUNCG

SUNCG数据集由普林斯顿大学的研究团队创建，是一个包含超过4.5万种室内环境的合成3D场景数据集。该数据集通过Planner5D平台生成，涵盖了从单室公寓到多层住宅等多样化场景，包含77.5万多个房间和569万余个物体实例，覆盖84个类别。数据集的创建过程包括利用在线室内设计工具生成场景，并通过人工标注和筛选确保数据质量。SUNCG旨在为计算机视觉和机器人领域的3D场景理解、物体检测与语义分割等任务提供丰富的训练资源，助力模型学习场景结构和物体语义信息。

The SUNCG dataset, developed by a research team at Princeton University, is a synthetic 3D scene dataset containing over 45,000 indoor environments. Generated through the Planner5D platform, the dataset covers diverse scenarios ranging from single-room apartments to multi-story residences, with more than 775,000 rooms and 5.69 million object instances across 84 categories. The creation workflow of SUNCG entails generating scenes via online interior design tools, followed by manual annotation and filtering to guarantee data quality. Designed to offer rich training resources for tasks including 3D scene understanding, object detection, and semantic segmentation in the fields of computer vision and robotics, SUNCG assists models in learning scene structures and semantic information of objects.

搜集汇总

背景与挑战

背景概述

SUNCG是一个大型合成3D场景数据集，专为语义场景补全任务设计，包含人工创建的密集体积标注，用于从单视角深度图像中同时预测体素占用和语义标签。该数据集支持3D上下文学习，通过SSCNet模型在CVPR 2017论文中提出，并提供访问协议、网页工具和C++工具箱供研究人员使用。