2D-3D-S

该数据集是一个大规模的集合，它将2D图像和3D信息相结合，专为语义分割任务而设计。RGB图像通过使用训练集的统计数据进行归一化，以符合零均值高斯分布；而深度图像则转换为HHA编码，并在[0, 1]的范围内进行归一化。该数据集的规模为：训练集包含52,903个样本，测试集包含17,593个样本，类别数为13个。所涉及的任务是语义分割。

This large-scale dataset integrates 2D images and 3D information, specifically designed for semantic segmentation tasks. RGB images are normalized to have a zero-mean Gaussian distribution using the statistics from the training set; depth images are converted to HHA encoding and then normalized within the range of [0, 1]. The dataset has the following specifications: the training set consists of 52,903 samples, the test set consists of 17,593 samples, and there are 13 semantic categories in total. The task addressed by this dataset is semantic segmentation.

搜集汇总

背景与挑战

背景概述

2D-3D-S是一个多模态室内场景理解数据集，提供超过70000张RGB图像以及对应的深度、表面法线、语义标注和3D网格等数据，覆盖6个室内区域总面积超过6000平方米。该数据集旨在支持联合2D-3D语义分析，适用于开发跨模态学习模型和室内场景理解研究。