Perceptual metric for graphic meshes

Abstract

The use of 3D models to represent data is growing increasingly common in fields such as architecture and digital entertainment. 3D meshes are subject to numerous processing operations. These operations may introduce distortions on the 3D meshes, and deteriorate the visual quality of the data.

Perceptual metrics are used to predict the visual quality of a model perceived by a human observer, by comparing a distorted model to its corresponding undistorted reference. In this study, numerous features are extracted and we examine their ability to measure the difference in visual quality between two models. The features considered are Gaussian weighted average, standard deviation, covariance, histogram and entropy of the mean curvatures of the vertices of a 3D model. The performance of these features in quality evaluation is tested on two datasets of models which contain a number of models affected by noise and smoothing distortions. The best features are then used to develop a metric that predicts mesh quality in a way that correlates well with human evaluation of distorted models.