A no-reference point cloud quality assessment method using the sparse convolutional neural network.
python 3.7
cudatoolkit 11.1
pytorch 1.8
minkowskiengine 0.5
xlrd 1.2
open3d
easydict
numpy
scipy
tensorboardX
pandas
tqdm
The maximum edge of the bounding box for input point cloud is required to be normalized to a uniform scale.
To reproduce the experiment results in the paper, the xyz coordinates of training samples are required to be normalized into 0-2000 (or 11-bit) for the voxel size setting of 5.
train.py is the main program. The explanations of some parameters in config.py are as below:
trainer_arg.add_argument('--train_file', type=str, default='./config/train.xlsx', help='file name and MOS for training set')
trainer_arg.add_argument('--test_file', type=str, default='./config/test.xlsx', help='file name and MOS for testing set')
trainer_arg.add_argument('--train_path', type=str, default='./config/path.xlsx', help='file name and file path for training set')
trainer_arg.add_argument('--test_path', type=str, default='./config/path.xlsx', help='file name and file path for testing set')
To load the existing model, the related parameter in config.py is:
misc_arg.add_argument('--resume', type=str, default='checkpoints/test_checkpoint-LS.pth', help='path for loading the checkpoint')
ps: test_checkpoint-LS.pth is trained on LS-PCQA dataset.
We establish a large-scale point cloud quality assessment dataset named LS-PCQA includes 104 reference point clouds and more than 22,000 distorted samples. The details can be found in our website and our paper (Point Cloud Quality Assessment: Dataset Construction and Learning-based No-Reference Metric).
Link for reference point clouds: BaiduNetDisk OneDrive
930 distorted samples with accurate MOS are supplied. Link: BaiduNetDisk OneDrive
Link for whole dataset with generated pseudo MOS: BaiduNetDisk OneDrive
If you use this code or our dataset please cite the paper
"Yipeng Liu, Qi Yang, Yiling Xu, Le Yang, "Point Cloud Quality Assessment: Dataset Construction and Learning-based No-Reference Metric", ACM Transactions on Multimedia Computing Communications and Applications, 2022."
@article{Liu2022ResSCNN,
title={Point Cloud Quality Assessment: Dataset Construction and Learning-based No-Reference Metric},
author={Yipeng Liu and Qi Yang and Yiling Xu and Le Yang},
journal={ACM Transactions on Multimedia Computing Communications and Applications},
year={2022}
}