ScanObjectNN

ScanObjectNN是由新加坡国立大学发布的一个用于真实场景下三维点云物体分类任务的典型基准数据集 。与常见的合成数据集不同，ScanObjectNN专为评估点云分类算法在嘈杂环境下的鲁棒性而设计 。它是目前少有的直接从真实RGB-D室内场景扫描（如ScanNet）中分割并经过人工清洗与标注而成的点云数据集 。由于真实扫描数据中天然存在遮挡、背景干扰、扫描噪声以及姿态变换，该数据集能更贴近实际应用场景，对于研究三维视觉算法在现实世界中的泛化能力和鲁棒性学习具有重要的学术价值 。该数据集的数据来源于ScanNet等RGB-D室内场景扫描数据集 。研究人员利用Mask R-CNN等方法从复杂的室内场景点云中分割出独立的物体实例，并进行了后续的人工清洗与类别标注 。数据集主要记录了静态的三维物体点云数据，存储格式为.npy（NumPy数组）。每个样本包含点的三维坐标（x, y, z），部分版本还包含法线或颜色信息，且所有数据已被统一采样到固定数量的点（如1024或2048个点）以适配深度学习模型 。 为了全面评估算法性能，数据集提供了多个预处理子集，包括仅含物体的OBJ_ONLY版本、保留背景干扰的OBJ_BG版本，以及最具挑战性的PB_T50_RS版本（包含随机扰动、旋转、平移及背景噪声），后者通常作为默认的评估基准 。ScanObjectNN涵盖了15个常见的室内物体类别，具体包括椅子、桌子、门、垃圾桶、显示器等 。数据集共包含约2,902个独立的三维物体实例 。

ScanObjectNN is a typical benchmark dataset released by the National University of Singapore for 3D point cloud object classification tasks in real-world scenarios. Unlike common synthetic datasets, ScanObjectNN is specifically designed to evaluate the robustness of point cloud classification algorithms in noisy environments. It is one of the few point cloud datasets directly segmented, manually cleaned and annotated from real RGB-D indoor scene scans such as ScanNet. Owing to the inherent occlusion, background interference, scanning noise and pose variations present in real scanned data, this dataset can better align with real-world application scenarios, holding significant academic value for researching the generalization capability and robust learning of 3D vision algorithms in the physical world. The data of this dataset is sourced from RGB-D indoor scene scan datasets including ScanNet. Researchers utilized methods such as Mask R-CNN to segment independent object instances from complex indoor scene point clouds, followed by subsequent manual cleaning and category annotation. The dataset mainly records static 3D object point cloud data, stored in .npy (NumPy array) format. Each sample contains the 3D coordinates (x, y, z) of points; some versions also include normal or color information, and all data has been uniformly sampled to a fixed number of points (e.g., 1024 or 2048 points) to adapt to deep learning models. To comprehensively evaluate algorithm performance, the dataset provides multiple preprocessed subsets, including the OBJ_ONLY version containing only objects, the OBJ_BG version retaining background interference, and the most challenging PB_T50_RS version (including random perturbations, rotations, translations and background noise), which is typically used as the default evaluation benchmark. ScanObjectNN covers 15 common indoor object categories, specifically including chairs, tables, doors, trash cans, monitors and so on. The dataset contains a total of approximately 2,902 independent 3D object instances.

搜集汇总

背景与挑战

背景概述

ScanObjectNN是一个真实场景下的三维点云物体分类基准数据集，包含15类室内物体共2,902个实例，专为评估算法在嘈杂环境中的鲁棒性而设计。数据集提供多个预处理子集，包括OBJ_ONLY、OBJ_BG和最具挑战性的PB_T50_RS版本，适用于研究三维视觉算法在现实世界中的泛化能力。