3DOS

3DOS数据集是由都灵理工大学的Antonio Alliegro、Francesco Cappio Borlino和Tatiana Tommasi共同创建的，旨在为3D开放集学习提供一个广泛的测试平台。该数据集包含51,127个来自55个常见物体类别的合成实例，用于评估在不同语义（类别）转移难度下的性能。数据集分为三个主要轨道：合成、真实到真实和合成到真实，旨在模拟真实世界的部署条件。3DOS数据集的应用领域包括机器人和自动驾驶等安全关键应用，旨在解决模型在开放世界中遇到未知类别时的识别和拒绝问题。The 3DOS Dataset was co-created by Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi from Politecnico di Torino, aiming to provide a comprehensive testbed for 3D open-set learning. This dataset contains 51,127 synthetic instances belonging to 55 common object categories, which is used to evaluate model performance under varying degrees of semantic (category) shift difficulty. The dataset is divided into three main tracks: Synthetic, Real-to-Real, and Synthetic-to-Real, designed to simulate real-world deployment conditions. The 3DOS Dataset covers safety-critical application domains such as robotics and autonomous driving, aiming to solve the problem of unknown category recognition and rejection that models encounter in open-world scenarios.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于3D开放集学习的基准测试数据集，专注于点云数据的语义新颖性检测。它包含多个子集，支持合成到合成和合成到真实的跨域场景，旨在评估模型在未知类别下的性能。该数据集已被NeurIPS 2022接受，提供了完整的代码实现和实验复现指南。