GLiDR: Topologically Regularized Graph Generative Network for Sparse LiDAR Point Clouds

CVPR 2024

Prashant Kumar*, Kshitij Bhat, Vedang Nadkarni, Prem Kumar Kalra

News

• [Feb 2024] GLiDR accepted in CVPR'2024.

Installation

The code was tested using pytorch 3.9.18. For topology based installation requirement we use the this fantastic TopologyLayer installation. It mostly works out of the box. There may be some issues while installing Dionysus Drivers for persistence calculation. Pleae refer to the issues at the above repository for quick resolution.

Alternatively you may also follow the below conda environment setup.

git clone https://github.com/prashkmr/GLiDR
cd GLiDR
conda env create -f env.yml
conda activate GLiDR

Dataset Setup

For training GLiDR we require paired correspondences for static and dynamic scans. These are available for CARLA and ARD-16. For KITTI, we generate the paired correspondence using a novel method that we develop in our paper. See our Supplementary material for precise details.

• [CARLA] - The CARLA dataset consists of 15 correspondence numpy arrays (0..7 for training and the rest for testing) for training and testing the model. We use the first three numpy arrays(0,1,2) to train the model (due to time intensive persistence computation). We further test the model on 8 (8,9...15) numpy arrays. For SLAM we use the 4 CARLA SLAM sequences made available. The data is available here. Download them in the data/carla folder and split into 2 subfolders - static and dynamic.
• [KITTI] - For KITTI we use the KITTI Odometry sequences for training the model. We transform the 11 sequences into numpy files and make them available here. We use the sequence 8 for testing the model and the rest for training our model For SLAM we use all the Odometry sequences (0 to 10). The static-dynamic corresponding pairs for KITTI are available here. Download them in the data/kitti folder and split into 2 subfolders - static and dynamic.
• [ARD-16] - The ARD-16 dataset consists of 4 numpy arrays (0..3). We use three (0..2) numpy arrays for training our model and the fourth numpy array for testing our model. The data is available here. Download them in the data/ard folder and split into 2 subfolders - static and dynamic.

Training

We trained our models on a single NVIDIA A100 GPU. We have 2 version per dataset - sparse and dense for training. For KITTI and CARLA sparse and dense version consists of 16 and 64 beam respectively. For KITTI the sparse and dense version consists of 8 and 16 beams respectively.

1. Train on KITTI dataset:

The training requires 2 arguments. --beam : Denotes the number of beam that ar allowed in the LiDAR. --dim : Sparsifies the outermost dimension of the range image (for KITTI and ARD-16, outermost dimesion is 1024, for CARLA it is 512). For more details on this, please refer to Section 5.2 of the paper.

Sparse Version

cd kitti/

python GLiDR_kitti.py --data data/kitti/ --exp_name glidr_kitti_sparse --beam 16 --dim 8 --batch_size 32 --mode kitti

Dense Version

cd kitti/
python GLiDR_kitti.py --data data/kitti/ --exp_name glidr_kitti_dense --beam 64 --dim 8 --batch_size 8 --mode kitti

2. Train on CARLA dataset:

Sparse Version

cd carla/
python GLiDR_kitti.py --data data/carla/ --exp_name glidr_carla_sparse --beam 16 --dim 4 --batch_size 32 --mode carla

Dense Version

cd carla/
python GLiDR_kitti.py --data data/kitti/ --exp_name glidr_carla_dense --beam 64 --dim 4 --batch_size 8 --mode carla

3. Train on CARLA dataset:
   Sparse Version

cd carla/
python GLiDR_kitti.py --data data/carla/ --exp_name glidr_carla_sparse --beam 16 --dim 4 --batch_size 32 --mode carla

Dense Version

cd carla/
python GLiDR_kitti.py --data data/kitti/ --exp_name glidr_carla_dense --beam 64 --dim 4 --batch_size 8 --mode carla

Contact

If you have any questions about our code or paper, kindly raise an issue on this repository.

Acknowledgment

TBD

BibTeX (Citation)

If you find our work useful in your research, please consider citing using:

@article{kumar2023glidr,
  title={GLiDR: Topologically Regularized Graph Generative Network for Sparse LiDAR Point Clouds},
  author={Kumar, Prashant and Bhat, Kshitij Madhav and Nadkarni, Vedang Bhupesh Shenvi and Kalra, Prem},
  journal={arXiv preprint arXiv:2312.00068},
  year={2023}
}