Learning to Generate Realistic LiDAR Point Clouds (LiDARGen) Paper

This repository contains the official implementation of our paper Learning to Generate Realistic LiDAR Point Clouds.

Usage

Environment

Install all python packages for training and evaluation with conda environment setup file:

conda env create -f environment.yml
conda activate lidar3

Set up KITTI-360

Download KITTI-360 from http://www.cvlibs.net/datasets/kitti-360/ (only the 3D LiDAR readings are required)
Set the KITTI360_DATASET environment variable to the KITTI360 dataset path. export KITTI360_DATASET=/path/to/dataset/KITTI-360/.

Sampling from Pretrained KITTI-360 model

Clone this repository and navigate to the project root directory
Download the pretrained model: curl -L https://uofi.box.com/shared/static/ahnc453qpx6pa8o7ikt2rllr3cavwcob --output kitti_pretrained.tar.gz
Extract the model: tar -xzvf kitti_pretrained.tar.gz
Setup conda environment: conda env create --name lidargen --file=environment.yml
Sample with python lidargen.py --sample --exp kitti_pretrained --config kitti.yml

Metrics

FID

Running FID requires downloading the 1024 backbone file for rangenet++ the following link and extracting it to the folder rangenetpp/lidar_bonnetal_master/darknet53-1024. This model is provided by the RangeNet++ repository. Finally, lidargen needs to be run with the --fid option. This can be done by running the following commands:

curl -L http://www.ipb.uni-bonn.de/html/projects/bonnetal/lidar/semantic/models/darknet53-1024.tar.gz --output darknet53-1024.tar.gz
tar -xzvf darknet53-1024.tar.gz
mv darknet53-1024/backbone rangenetpp/lidar_bonnetal_master/darknet53-1024/
rm darknet53-1024.tar.gz
rm -r darknet53-1024
python lidargen.py --fid --exp kitti_pretrained --config kitti.yml

MMD

python lidargen.py --mmd --exp kitti_pretrained --config kitti.yml

JSD

python lidargen.py --jsd --exp kitti_pretrained --config kitti.yml

Denoising Visualization

To create a denoising visualization, LiDARGen needs to be run manually.

cd LiDARGen
curl -L https://uofi.box.com/shared/static/ahnc453qpx6pa8o7ikt2rllr3cavwcob --output kitti_pretrained.tar.gz
tar -xzvf kitti_pretrained.tar.gz
conda activate lidargen
python main.py --config kitti.yml --exp kitti_pretrained --sample --ni
python visualization.py --exp kitti_pretrained

A video similar to the following will be produced:

Densification MAE and IOU

To perform densification run:

python lidargen.py --densification --exp kitti_pretrained --config kitti.yml

The target range images will be outputted to kitti_pretrained/densification_target. The LiDARGen densification result images will be outputted to kitti_pretrained/densification_result.

To calculate MAE (Mean Average Error) run:

python lidargen.py --mae --exp kitti_pretrained --config kitti.yml

To calculate IOU run:

python lidargen.py --iou --exp kitti_pretrained --config kitti.yml

Samples

14k Samples

We provide a collection of 14,375 samples for research purposes here: 14K LiDARGen Samples.

Baselines and ECCV'22 Benchmark Samples

We provide the LiDARGen and baseline samples used for performing the FID evaluation in our paper. The FID scores have marginal differences from the paper due to the random sampling performed to make FID calculation for RangeNet++ features possible.

For any new applications we recommend using the LiDARGen samples from the "20k Samples" section above (which were generated with a newer model).

Method	FID Score
pc_ncsn	2116.1611
pc_projectedgan	2188.4207
pc_gan	3016.7556
pc_vae	2298.7101

Samples are available at this link. Instructions to generate FID scores follow (replace the string "pc_projectedgan" in each instruction with the folder you chose):

curl -L https://uofi.box.com/shared/static/kmfe6alnt9xvnxnsvmv648a5mo8lq0j5.gz --output lidargen_eccv22_samples.tar.gz
tar -xzf lidargen_eccv22_samples.tar.gz
python lidargen.py --fid_pointcloud pc_projectedgan --folder_name pc_projectedgan
python lidargen.py --fid --fid_folder1 pc_projectedgan_fid

Additional Information

Range Image .pth format

LiDARGen represents LiDAR readings in a range image format (i.e., a 360 degree depth map with an additional intensity layer). These range images are normalized into a [0, 1] range with the transform range_normalized = (np.log2(range_unnormalized+1)) / 6.

When sampling or densification is performed, LiDARGen generates .pth files. These are serialized PyTorch tensors which can be loaded with range_im = torch.load('filename.pth'). The shape of this tensor is typically [2, 64, 1024] where the dimension represent [channel, height, width]. The range image depth value can be converted into a meter format by running range_unnormalized = torch.exp2(range_im[0]*6) - 1

Bibtex

@inproceedings{zyrianov2022learning,
  title={Learning to Generate Realistic LiDAR Point Cloud},
  author={Zyrianov, Vlas and Zhu, Xiyue and Wang, Shenlong},
  booktitle={ECCV},
  year={2022}
}

raven38/lidargen