Predição de descritores geométricos em nuvens de pontos utilizando a arquitetura PointNet

Abstract

Geometric descriptors are characteristics or mathematical representations that capture information about the geometry of objects or scenes and are widely used in matching points between different point clouds in the context of registration. Geometric descriptors are often computatio nally expensive to calculate due to the lack of ordering in the structure of point clouds, requiring iterative neighbor search for each point. Therefore, it is desirable to optimize the determination of local geometric descriptors. In this work, five architectures with three different strategies for point-to-point regression based on PointNet were proposed. Although the developed models did not achieve the goal of creating an approximating function for calculating geometric descriptors, the modeling and testing process yielded important insights regarding the use of neural networks in point clouds.