2D-3D-S (2D-3D-Semantic)

2D-3D-S数据集提供了来自2D，2.5D和3D域的各种相互注册的模态，并带有实例级语义和几何注释。它覆盖了6个大型室内区域收集的6,000平方米多个，这些区域起源于3个不同的建筑。它包含70,000多个RGB图像，以及相应的深度，表面法线，语义注释，全局XYZ图像 (所有形式的规则和360 ° 等矩形图像) 以及相机信息。它还包括注册的原始和语义注释的3D网格和点云。该数据集可以利用大型室内空间中存在的规律性来开发联合和跨模式学习模型以及潜在的无监督方法。The 2D-3D-S dataset provides various mutually registered modalities from the 2D, 2.5D, and 3D domains, paired with instance-level semantic and geometric annotations. It spans over 6,000 square meters across 6 large indoor areas collected from 3 distinct buildings. The dataset contains more than 70,000 RGB images, along with corresponding depth maps, surface normals, semantic annotations, global XYZ images (covering both regular and 360° equirectangular formats), and camera metadata. It also includes registered raw and semantically annotated 3D meshes and point clouds. This dataset can leverage the inherent regularities present in large-scale indoor spaces to develop joint and cross-modal learning models as well as potential unsupervised approaches.

搜集汇总

背景与挑战

背景概述

2D-3D-S数据集是一个大型室内场景数据集，提供2D、2.5D和3D域的多模态注册数据，包括超过70,000张RGB图像及深度、语义注释、点云等，覆盖6个区域超过6,000平方米。其特点是包含实例级语义和几何注释，旨在支持联合跨模态学习模型和三维语义分割等计算机视觉任务研究。