TransCG

TransCG 是第一个用于透明对象深度补全和抓取的大型真实世界数据集，其中包含 51 个透明对象的 57,715 个 RGB-D 图像以及从真实世界下 130 个场景的不同视角（~240 个视点）捕获的许多不透明对象设置。样本由两种不同类型的相机（Realsense D435 和 L515）捕获。提供以下数据： 透明物体的3D模型；每个场景的每个视点中透明物体的6dpose；原始 RGB-D 图像和地面实况细化深度图像；透明物体的蒙版；每个样本的真实表面法线。TransCG is the first large-scale real-world dataset dedicated to transparent object depth completion and robotic grasping. It contains 57,715 RGB-D images corresponding to 51 distinct transparent objects, alongside numerous cluttered opaque object setups captured from approximately 240 viewpoints across 130 real-world scenes. All samples are collected using two types of cameras: Intel RealSense D435 and L515. The following data is provided: 3D models of transparent objects; 6D poses of transparent objects at each viewpoint for every scene; raw RGB-D images and ground-truth refined depth images; masks of transparent objects; and ground-truth surface normals for each sample.

搜集汇总

背景与挑战

背景概述

TransCG是首个用于透明对象深度补全和抓取的大型真实世界数据集，包含57,715个RGB-D图像，覆盖51个透明对象和130个场景的多个视角。该数据集提供丰富的标注信息，如3D模型、6D姿态、地面实况深度图像和表面法线，旨在支持透明物体深度估计和机器人抓取研究。