We propose the first robustness benchmark of point cloud detectors against common corruption patterns. We first introduce different corruption patterns collected for this benchmark and dataset. Then we propose the evaluation metrics used in our benchmark. Finally, we introduce the subject object detection methods and robustness enhancment methods selected for this benchmark.

All the codes are tested (not only) on Ubuntu 20.04 with Python 3.8 tools:

• PyGeM and PyMieScatt
• Scipy, and other basic tools, like numpy, os, argparse, glob, etc.
• As for the weather simulation, please refer to LISA and Lidar_fog_simulation.

Most tools can be installed by pip install ${tool}.

This kit of corruption simulation applies to KITTI data. Please download the KITTI 3D object detection dataset and organize the files as the below:

data
├── kitti
│   │── ImageSets
│   │── training
│   │   ├──calib & velodyne & label_2 & image_2 
│   │── testing (optional)
│   │   ├──calib & velodyne & image_2

More details of the implementation are as in object and scene folder.

We formulate 25 corruptions covering 2 affecting ranges $\times$ 4 corruption categories, i.e., {object, scene} $\times$ {weather, noise, density, transformation}.

For the implementation of 25 corruptions, corresponding README files under the object and scene folders give the details on the Python commands to generate simulated data.

We show some corruption examples based on the KITTI LiDAR example with ID = 000008, as in the below figures. Besides, we provide the ground-truth annotations of objects and detection results obtained by PVRCNN in the format of bounding boxes.

As in the paper, we tested 7 voxel-based methods, 3 point-voxel-based methods, and 2 point-based methods (we re-shaped PartA2 with point-based data representation and pointnet++ backbone) on the OpenPCDet (https://github.com/open-mmlab/OpenPCDet). Also, we extended CenterPoint to the versions of different data representations and different proposal architectures for relatively fair evaluation.

Some parts of the code implement are learned from the official released codes of the below methods:

• Snow and Rain simulation: LISA
• Fog simulation: Lidar_fog_simulation
• Object-level corruptions: ModelNet40-C
• Extraction of object-level points: OpenPCDet

We would like to thank for their proposed methods and the official implementation.

If you find this project useful in your research, please consider cite:

@article{li2022common,
  title={Common Corruption Robustness of Point Cloud Detectors: Benchmark and Enhancement},
  author={Li, Shuangzhi and Wang, Zhijie and Juefei-Xu, Felix and Guo, Qing and Li, Xingyu and Ma, Lei},
  journal={arXiv preprint arXiv:2210.05896},
  year={2022}
}