BitDance PCQA Metric

BitDance: Low Complexity Point Cloud Quality Assessment Metric

BitDance is a fast, OpenMP optimized implementation of the technique described in Color and Geometry Texture Descriptors for Point-Cloud Quality Assessment. If you find this code useful in your research, please consider citing:

@ARTICLE{9450013,
   author={Diniz, Rafael and Freitas, Pedro Garcia and Farias, Mylène C. Q.},
   journal={IEEE Signal Processing Letters}, 
   title={Color and Geometry Texture Descriptors for Point-Cloud Quality Assessment}, 
   year={2021},
   volume={28},
   number={},
   pages={1150-1154},
   doi={10.1109/LSP.2021.3088059}}

This repository contains the code for all discussed algorithms in the article. The C++ code produces three binaries: bitdance_pcqa, create_normals and optimize_voxel_size. The bitdance_pcqa calculates the color and geometry PC feature statistics; create_normals creates PCs with oriented normals; and optimize_voxel_size uses a parameterized heuristic and constant to provide a voxel size for a given PC. In directory "distance_calculation", it is present the python script we developed for statistics distance calculation, which includes the Jensen-Shannon divergence code, among others.

Compilation

This code depends on the Open3D library: https://github.com/intel-isl/Open3D

The code was compiled and tested on Debian Linux 11 (codename Bullseye) and Ubuntu 22.04 LTS. To install the Open3D library and other dependencies, run (as root):

apt-get install libopen3d-dev libeigen3-dev libpng-dev  zlib1g-dev

In order to compile the code, written in C++, open the Makefile, adjust it to your preferred C++ compiler, set to your Open3D installation prefix path (no need for any modification on Debian 11 or Ubuntu 22.04), and run:

make

Usage

Prior to the execution of the metric, normals without ambiguities need to be created. For this use:

create_normals point_cloud.ply pc_with_normals.ply

In order to obtain the target voxel edge size of a PC, given a "k" constant ("k" equals to 6.0 in the article), using the average distance of 8 nearest neighbors (hardcoded in the code) as reference, as described in the article, use:

optimize_voxel_size 3 k point_cloud.ply

To run the CIEDE2000 color-based feature extractor with 8 bits label, neighborhood size of 12, voxel size (use the output of optimize_voxel_size) "E", with the statistics (histograms) results written to "results-c.csv", use:

bitdance_pcqa -i point_cloud.ply -n 12 -m 0,1,0,0,0 -v E -h results-c.csv

To run the geometry-based feature extractor with label of 16-bit, neighborhood size of 6, without the use of voxelization (voxelization does not help for the geometry-based feature extractor), with the statistics (histograms) results written to "results-g.csv", use:

bitdance_pcqa -i pc_with_normals.ply -n 6 -m 0,0,1,0,0 -h results-g.csv

Authors

• Rafael Diniz (rafael@riseup.net)

• Pedro Garcia Freitas (pedro@sawp.com.br) with some metrics calculation python code

Acknowledgments

• Mylene C.Q. Farias, my PhD advisor

• Pedro Garcia Freitas, my PhD co-advisor

• Open3D by Qian-Yi Zhou and Jaesik Park and Vladlen Koltun: http://open3d.org/

• Color spaces conversion code by Nicolae Berendea: https://github.com/berendeanicolae/ColorSpace

• Financial support by the following Brazilian agencies: CAPES and FAP-DF

License

This code is licensed under GNU General Public License version 3 or any higher version, while the color spaces conversion and the Open3D code are MIT licensed.