2D-3D Match Dataset

2D-3D 匹配数据集是一个新的 2D-3D 对应数据集，它利用了来自 RGB-D 扫描的多个 3D 数据集的可用性。具体来说，使用了来自 SceneNN 和 3DMatch 的数据。训练数据集由 110 个 RGB-D 扫描组成，其中 56 个场景来自 SceneNN，54 个场景来自 3DMatch。如下生成2D-3D对应数据。给定从 3D 点云中随机采样的 3D 点，从不同的扫描视图中提取一组 3D 补丁。为了找到 2D-3D 对应关系，对于每个 3D 补丁，其 3D 位置被重新投影到该点位于相机平截头体中的所有 RGB-D 帧中，同时考虑到遮挡。提取重新投影点周围的相应局部 2D 补丁。总共收集了大约 140 万个 2D-3D 对应关系。

The 2D-3D matching dataset is a novel 2D-3D correspondence dataset that leverages the availability of multiple 3D datasets acquired from RGB-D scans. Specifically, data from SceneNN and 3DMatch are utilized for this dataset. The training dataset consists of 110 RGB-D scans, where 56 scenes are sourced from SceneNN and 54 scenes from 3DMatch. The 2D-3D correspondence data is generated as follows: Given 3D points randomly sampled from a 3D point cloud, a set of 3D patches are extracted from different scan views. To establish 2D-3D correspondences, the 3D position of each 3D patch is reprojected onto all RGB-D frames in which the point lies within the camera frustum, while accounting for occlusion. Corresponding local 2D patches around the reprojected points are then extracted. In total, approximately 1.4 million 2D-3D correspondences are collected.

搜集汇总

背景与挑战

背景概述

2D-3D Match Dataset是一个用于2D-3D匹配任务的数据集，基于SceneNN和3DMatch的RGB-D扫描构建，包含110个训练场景并生成了约140万个2D-3D对应关系，旨在支持跨域描述符学习研究。