This is the official code for our paper "SCP: Spherical-Coordinate-based Learned Point Cloud Compression". paper
Directly run the install.sh script:
./install.sh
Please refer to the official website of KITTI, Then download the 80GB dataset "odometry data set (velodyne laser data, 80 GB)".
After getting the KITTI dataset, unzip the file in data/kitti/.
You need to be a member of MPEG first. The Ford dataset is provided on their website.
It is better to use multiple processes to preprocess the data. It can be much longer time if you use only one process. We provide a data_preproc/data_preprocess.py, whose usage is:
python data_preproc/data_preprocess.py <proc_number> <command>
as used below.
Generating training set. Note: the generated dataset is big (KITTI dataset ~2TB, Ford is much smaller), take care of your storage.
# Spherical KITTI
python data_preproc/multi_data_preproc.py 128 python data_preproc/data_preprocess.py --type kitti --ori_dir data/kitti/sequences/\*/velodyne/\*.bin --out_dir data/kitti/spher --spher
# Spherical Ford
python data_preproc/multi_data_preproc.py 128 python data_preproc/data_preprocess.py --type ford --ori_dir data/ford/Ford_01_q_1mm/\*.ply --out_dir data/ford/spher --spher
You can remove --spher to generate Cartesian-coordinate-based data, remember to change the --out_dir.
If you do not need D2 PSNR on KITTI dataset, you can skip this step.
Calculate normals for KITTI dataset to calculate D2 PSNR.
python data_preproc/multi_data_preproc.py 128 python data_preproc/gene_normals.py --ori_dir data/kitti/sequences/test/\*/velodyne/\*.bin --out_dir data/kitti/test_norm
Generate testing data with multi-level Octree.
You can change the lidar_level from 11 to 16 for KITTI and 11 to 17 for Ford. Note: the generated dataset is big, take care of your storage. You can add --spher or --cylin to generate spherical/cylindrical-coordinate-based data. You can further add --mullevel to generate multi-level octree.
python data_preproc/multi_data_preproc.py 128 python data_preproc/test_gene.py --type kitti --ori_dir data/kitti/test_norm/\*/\*.ply --out_dir data/kitti/spher_mullevel_16 --lidar_level 16
python data_preproc/multi_data_preproc.py 128 python data_preproc/test_gene.py --type ford --ori_dir data/ford/test/\*/\*.ply --out_dir data/ford/mullevel_17 --lidar_level 17
Train the EHEM model with generated Spherical-coordinate-based data.
# SCP-EHEM KITTI
python train.py --config-dir configs/ --config-name train_kitti_ehem.yaml data.batch_size=16 gpus=\[0,1\] data.root=data/kitti/spher/\*.npy
# SCP-EHEM Ford
python train.py --config-dir configs/ --config-name train_ford_ehem.yaml data.batch_size=16 gpus=\[0,1\] data.root=data/ford/spher/\*.npy
Evaluate the SCP-EHEM model with generated testing data. Following are some examples:
CUDA_VISIBLE_DEVICES=0 python encode_mullevel.py --type kitti --lidar_level 16 --ckpt_path outputs/kitti/2023-07-05/15-21-55/ckpt/epoch=19-step=124840.ckpt --spher --preproc_path data/kitti/mullevel_16/ --test_files data/kitti/test_norm/17
CUDA_VISIBLE_DEVICES=0 python encode.py --type kitti --lidar_level 12 --ckpt_path outputs/kitti/2023-07-05/15-21-55/ckpt/epoch=19-step=124840.ckpt --spher --preproc_path data/kitti/samelevel_12/ --test_files data/kitti/test_norm/17
CUDA_VISIBLE_DEVICES=2 python encode.py --type kitti --lidar_level 14 --ckpt_path outputs/kitti/2023-07-04/17-21-18/ckpt/epoch=8-step=93303.ckpt --cylin --preproc_path data/kitti/cylinlevel_14/ --test_files data/kitti/test_norm/17
The preproc_path is generated in "Testing data preproc".
SCP-EHEM and SCP-OctAttention for KITTI and Ford dataset: link
If you found this work is helpful for you, please cite:
@inproceedings{luo2024scp,
title={SCP: Spherical-Coordinate-Based Learned Point Cloud Compression},
author={Luo, Ao and Song, Linxin and Nonaka, Keisuke and Unno, Kyohei and Sun, Heming and Goto, Masayuki and Katto, Jiro},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={4},
pages={3954--3962},
year={2024}
}