RangeLDM

[ECCV 2024] Official implementation of "RangeLDM: Fast Realistic LiDAR Point Cloud Generation"

Models

KITTI360

Model	MMD	FRD	JSD	Checkpoint	Generated Point Clouds
RangeLDM	3.07 × 10^−5	1074.9	0.045	[PKU Disk] (115MB)	[1k samples]
RangeDM	4.14 × 10^−5	899.0	0.040	[PKU Disk] (401MB)	[1k samples]

nuScenes

Model	MMD	JSD	Checkpoint	Generated Point Clouds
RangeLDM	1.9 × 10^−4	0.054	[PKU Disk] (153MB)	[1k samples]

Train

VAE

cd vae
python main.py --base configs/kitti360.yaml

LDM

cd ldm
accelerate launch train_unconditional.py --cfg configs/RangeLDM.yaml # for unconditional generation
accelerate launch train_conditional.py --cfg configs/upsample.yaml # for conditional generation

Evaluation

see metrics/metrics.md

Citation

If you find our work useful, please cite:

@article{hu2024rangeldm,
  title={RangeLDM: Fast Realistic LiDAR Point Cloud Generation},
  author={Hu, Qianjiang and Zhang, Zhimin and Hu, Wei},
  journal={arXiv preprint arXiv:2403.10094},
  year={2024}
}