Data-driven point sampling with blue-noise properties for triangular meshes.

In this paper, we present a novel data-driven schema for generating stochastically distributed points with blue-noise properties on triangle meshes. We propose a pre-processing method to generate triangles with different features and sample each of them uniformly. Subsequently, we utilize the k-nearest neighbors (k-NN) regression to train the features and the pre-sampled points of these triangles. For the inference of new meshes, we generate a point set with the Poisson distribution and minimal distance constraint based on the pre-sampled points of new meshes inferred by the trained model. The result of k-fold cross-validation shows that the trained model shows high-quality results dues to its small MSE. The frequency analysis shows that the sampled points on new meshes have blue-noise characterization. We also apply the trained model to sample the surfaces of 3D models. Result analysis shows that the proposed approach gives better performance than the state-of-art computational geometry methods.