kexuanxia/occam

Demo code for the paper: OccAM's Laser

OccAM's Laser

This is the demo code for the paper:

OccAM's Laser: Occlusion-based Attribution Maps for 3D Object Detectors on LiDAR Data
David Schinagl, Georg Krispel, Horst Possegger, Peter M. Roth, and Horst Bischof
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022
[Paper] [Supp.] [arXiv]

While 3D object detection in LiDAR point clouds is well-established in academia and industry, the explainability of these models is a largely unexplored field. In this paper, we propose a method to generate attribution maps for the detected objects in order to better understand the behavior of such models. These maps indicate the importance of each 3D point in predicting the specific objects. Our method works with black-box models: We do not require any prior knowledge of the architecture nor access to the model's internals, like parameters, activations or gradients. Our efficient perturbation-based approach empirically estimates the importance of each point by testing the model with randomly generated subsets of the input point cloud. Our sub-sampling strategy takes into account the special characteristics of LiDAR data, such as the depth-dependent point density.

This repository is based on [OpenPCDet] (version v0.5.2)

Overview

• Requirements
• Installation
• Demo Example
• Acknowledgement
• Citation

Requirements

• Linux (tested on Ubuntu 20.04 LTS)
• Python 3.8+
• CUDA (tested with version 11.3)
• PyTorch (tested with version 1.10)
• spconv 2.1.x (tested with version 2.1.21)

Installation

1) Clone this repository

git clone https://github.com/dschinagl/occam.git

2) Install PyTorch and SpConv depending on your environment

• PyTorch
• SpConv 2.1.x

3) Install required packages

pip install -r requirements.txt

4) Install pcdet

python setup.py develop

Demo Example

We provide a quick demo to create attribution maps for detections of a KITTI pretrained PointPillars model on custom point cloud data and to visualize the resulting maps.

1. Download the pretrained KITTI PointPillars model from the OpenPCDet model-zoo.

2. Prepare your custom point cloud data (skip this step if you use the provided sample demo_pcl.npy):

   • You need to transform the coordinate of your custom point cloud to the unified normative coordinate of OpenPCDet (x-axis front,y-axis left and z-axis top).
   • The origin of your coordinate system should be about 1.6m above the ground, since the provided model is trained on the KITTI dataset.
   • The point cloud shape should be (N, 4) -> [x, y, z, intensity]

3. Run the demo as follows:

python occam_demo.py --ckpt ${PRETRAINED_POINTPILLARS_MODEL} \
    --source_file_path ${POINT_CLOUD_DATA}

${POINT_CLOUD_DATA} could be:

• The provided sample demo_pcl.npy
• A numpy array file (N, 4) like my_data.npy
• Original KITTI .bin data like data/kitti/training/velodyne/000008.bin

Acknowledgement

We thank the authors of OpenPCDet for their open source release of their codebase.

Citation

If you find this code useful for your research, please cite

@InProceedings{Schinagl_2022_CVPR,
    author    = {Schinagl, David and Krispel, Georg and Possegger, Horst and Roth, Peter M. and Bischof, Horst},
    title     = {OccAM's Laser: Occlusion-Based Attribution Maps for 3D Object Detectors on LiDAR Data},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {1141-1150}
}