No-Reference Point Cloud Quality Assessment

Abstract

The need to capture environments and objects in 3 dimensions to produce a high quality digital representation is proving to be useful in many applications in the world, where there is an increasing dependence on digital spaces. Point Clouds are a data type to represent 3D objects and scenes. During the processing of the point cloud data, undesired changes to the point cloud can be introduced or information can get lost. Automatic detection and classifying of the quality of a point cloud is needed, to regulate this data type for use in 3D environments, and speed up the processing of point cloud objects in real-time applications.

In this thesis, we look into the problem of no-reference point cloud quality assessment. Full-reference quality metrics compare a distorted point cloud with its original, while no-reference or blind metrics only look at a single point cloud. This task is more challenging than full-reference, as there is less direct information available to work with. We approach the challenge of how the perceptual quality of a point cloud can be estimated when a high-quality reference point cloud is not available, by analyzing if any point cloud properties correlate with the perceptual quality.

We try to find point cloud properties that are effective to use in a model to predict a user given quality score. We select multiple local properties to transform into global descriptors, by taking multiple statistical attributes of the data on all the points to combine into scalars. We analyze the created global descriptors to find linear correlations with the quality score, and found many features that to some measure linearly correlate with some distortion types. With the selected global descriptors, we use Support Vector Regression to find the fitting of the features to the quality score. Varying performances are achieved by training the models on different data subsets, and different fitting kernel functions.

The results show some distortions on unknown objects are properly predictable using fitted weights on this distortion. However, a single defining quality descriptor is not yet found. While the results show potential, our model is not robust and fast enough to perform a reliable assessment of quality in a real-time environment. We are just at the beginning of exploring no-reference quality assessment, and while the current methods are not yet appliable in real-time scenarios, future work can fill many knowledge gaps to reach this goal.