ScanRefer Dataset

该数据集名为ScanRefer，包含了800个ScanNet场景中的11,046个物体的51,583条描述，这些描述不仅涉及物体的外观，还包括空间关系。该数据集被划分为训练集和验证集，分别包含36,665和9,508个样本，而测试集尚未正式发布。在规模上，该数据集为11,046个物体提供了51,583条描述，其任务重点是3D密集标注。

The dataset named ScanRefer contains 51,583 textual descriptions for 11,046 objects across 800 ScanNet scenes. These descriptions cover not only the visual appearance of the objects but also their spatial relationships. The dataset is split into training and validation subsets, which contain 36,665 and 9,508 samples respectively, while the test subset has not been officially released. Overall, this dataset provides 51,583 descriptions for 11,046 objects, with its core task focusing on dense 3D annotation.

搜集汇总

背景与挑战

背景概述

ScanRefer Dataset是一个用于3D场景密集标注任务的数据集，通过多阶关系挖掘技术提高标注准确性。该数据集支持多种点特征输入，并提供详细的训练和评估脚本，适用于计算机视觉领域的研究。