Comparative Analysis of Downsampling Techniques for Machine Learning on Cultural Heritage Objects

Abstract

Presenting point clouds in web environments is a very demanding process. It requires downsampling of point clouds in specific file formats, such us X3D, that uses web architecture to work across diverse devices and lightweight 3D graphics to ensure web page functionality. This paper evaluates the performance of several point cloud downsampling methods with the objective of identifying those that best balance data reduction for efficient web deployment with the preservation of essential features required for AI and machine learning applications. We present a comparative analysis of Voxel Grid Downsampling (VGD), Uniform Grid Subsampling (UGS), Curvature Preserving Sampling (CPS), Random Sampling (RS) and Farthest Point Sampling (FPS). To assess the impact of each approach on the distribution and structure of sampled point clouds, we conduct a number of experiments by combining SHREC 2021 Cultural Heritage dataset with PointNet.