Common Corruption Robustness Benchmark for Point Cloud Detectors
收藏arXiv2022-10-12 更新2024-08-06 收录
下载链接：
http://arxiv.org/abs/2210.05896v1
下载链接
链接失效反馈官方服务：
资源简介：
本研究构建了一个针对点云检测器的常见损坏鲁棒性基准，名为Common Corruption Robustness Benchmark。该数据集包含1,122,150个样本，覆盖了7,481个场景，涉及25种常见损坏类型和6种严重程度。数据集通过物理感知模拟方法生成，模拟了自然天气、噪声干扰、密度变化和物体变换等多种损坏情况。此数据集旨在为开发实用且鲁棒的点云检测器提供基础，通过广泛的实证研究评估现有检测器的脆弱性，并探索增强鲁棒性的方向。此外，研究还探讨了基于数据增强和数据去噪的现有鲁棒性增强方法的有效性及其局限性。
This study constructs a common corruption robustness benchmark for point cloud detectors, named Common Corruption Robustness Benchmark. This dataset contains 1,122,150 samples, covers 7,481 scenarios, and involves 25 common corruption types and 6 severity levels. The dataset is generated via physics-aware simulation methods, which simulates various corruptions such as natural weather conditions, noise interference, density variations, and object transformations. This dataset aims to provide a foundation for developing practical and robust point cloud detectors, evaluate the vulnerability of existing detectors through extensive empirical studies, and explore directions for enhancing robustness. Furthermore, this study also discusses the effectiveness and limitations of existing robustness enhancement methods based on data augmentation and data denoising.
提供机构：
阿尔伯塔大学, 阿尔伯塔机器智能研究所, 九州大学
创建时间：
2022-10-12
