nuScenes-UNION-labels
收藏资源简介:
nuScenes-UNION标签数据集是基于nuScenes数据集的,包含了使用UNION方法生成的预建的.pkl标注文件。UNION是一种无监督的3D目标检测方法,利用基于外观的伪类进行检测。该数据集分为三个顶层文件夹:标准标签、类无关标签和多类003标签,分别用于标准数据集划分和类别的MMDetection3D信息、类无关和多功能(车辆、行人、骑行者)3D目标检测的地面真实标签和伪标签。
The nuScenes-UNION annotation dataset is based on the nuScenes dataset, and contains pre-built .pkl annotation files generated via the UNION method. UNION is an unsupervised 3D object detection approach that leverages appearance-based pseudo-classes for detection. This dataset is split into three top-level folders: standard annotations, class-agnostic annotations, and multi-class 003 annotations. These three folders serve the following respective purposes: the standard annotations folder is used for standard dataset partitioning and category-specific MMDetection3D-related information; the class-agnostic annotations folder is used for class-agnostic 3D object detection ground-truth labels and pseudo-labels; and the multi-class 003 annotations folder is used for multi-functional (vehicle, pedestrian, cyclist) 3D object detection ground-truth labels and pseudo-labels.
nuScenes-UNION-labels 数据集概述
基本信息
- 数据集名称: nuScenes-UNION-labels
- 许可证: CC BY-NC-SA 4.0(非商业性使用,相同方式共享)
- 任务类别: 目标检测
- 语言: 英语
- 相关论文: UNION: Unsupervised 3D Object Detection using Appearance-based Pseudo-Classes(NeurIPS 2024)
数据集内容
- 标注类型: 伪标签(pseudo-labels)
- 数据来源: 基于nuScenes数据集生成
- 标注格式: MMDetection3D信息文件(.pkl格式)
- 文件夹结构:
standard-labels: 标准数据集分割和类别的真实标注class-agnostic-labels: 类别无关的3D目标检测的真实标签和伪标签multi-class-003-labels: 多类别(车辆、行人、骑行者)的3D目标检测的真实标签和伪标签
用途特点
- 专为无监督3D目标检测设计
- 基于外观的伪类别生成
- 快速复现UNION结果,无需重新生成标签
- 仅包含标注文件,不包含传感器数据
相关资源
- UNION代码库: https://github.com/TedLentsch/UNION
- nuScenes数据集: https://www.nuscenes.org/nuscenes
- 原始论文: https://arxiv.org/abs/2405.15688
使用要求
- 必须遵守原始nuScenes数据集的许可协议
- 使用时需要引用nuScenes、MMDetection3D和UNION的相关文献
- 仅限非商业用途




