Point Cloud Quality
Point Cloud Quality Assessment (PCQA) has achieved remarkable attention in recent years and applied to different point cloud media formats [1, 2]. In the case of FR or Full-Reference, we need to have a reference point cloud to achieve a metric to evaluate the quality of the point cloud. Without a Reference point cloud, we have to use No-Reference or NR method to assess the image quality.
In the following article, the different terminologies regarding point cloud analysis, quality assessment categories, description metrics, dataset, etc are explained briefly.
What is Point Cloud?
When we have a set of data points in space in a 3D shape with specific data point coordinates of (X, Y, and Z), we have point cloud data [1]. It also may contain extra attributes including color, reflectance, width, length, density, etc. To have a point cloud one may need typical gadgets such as scanners, LiDAR Cameras, or software.
Point clouds are getting more attention thanks to the capability of 3D machines. There are various applications of point clouds such as Virtual Reality (VR), Augmented Reality (AR), Self-driving cars (LiDAR data), etc.
Distortion
Same as the simple 2D images, the point clouds also have distortion caused by the laser beams used in the related devices to build the dataset and depth camera. Moreover, the tabular structure of the dataset has distortion as well, including missing coordinates, etc [4].
Mean opinion score (MOS)
MOS is a metric to demonstrate the quality experience scaling with an integer number between two bands ([1, 5]), with the highest perceived quality as the highest rational number. To collect the MOS for each dataset, the presentation device should be the same for all audiences with the standard and identical resolution and zoom rate. In this process, the images will be shown to each participant for a specific time period and the given score by them in terms of the image quality will collect.
Point Cloud Quality Assessment (PCQA)
An assessment needs to be defined to have metrics for quality evaluation of point clouds and distorted point cloud optimization.
The PCQA include the following metrics type;
- Full-Reference (FR)
- Reduced-Reference (RR)
- No-Reference (NR)
It should note that the Full-Reference (FR) and Reduced-Reference (RR) approaches need a reference point to apply the calculations over this point such as the point-to-point method. But what if we do not have access to the reference points? If so, we need to continue with the NR-PCQA metrics.
Also, if you need to build or test the quality metrics, you need the particular built PCQA dataset with standard distortion.
Point Cloud Assessment Dataset
Designing learning algorithms and quality assessment metrics on point clouds need a proper large-scale point cloud dataset. One of the missing items in this field is the small amount of open source large-scale point cloud datasets with labels, references, and noises with high quality, and also a reasonable value of MOS.
OCTree
An ocTree is a tree data design in which each of the branches may have at least 8 leaves. The leaf is called octane and utilizes to store 3D points of the point cloud image. When we are dealing with high resolution and big 3D inputs in graphics, it is better to use the OCTree.
References
- [1] Qiuping Jiang,Wei Gao, ShiqiWang, et al. 2020. Blind Image Quality Measurement by Exploiting High-Order Statistics With Deep Dictionary Encoding Network. IEEE Transactions on Instrumentation and Measurement 69, 10 (2020), 7398–7410.
- [2] Qiuping Jiang, Feng Shao, Weisi Lin, et al. 2018. Optimizing Multistage Discriminative Dictionaries for Blind Image Quality Assessment. IEEE Transactions on Multimedia 20, 8 (2018), 2035–2048.
- [1] Wikipedia
- [2] Liu, Yipeng, et al. “Point cloud quality assessment: Dataset construction and learning-based no-reference metric.” ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) (2022).
