Construction and Visualization of Gaussian Mixture Models from 3D Point Clouds for Deep Learning (DAAV)

Gaussian Mixture Models (GMMs) are probability density functions based on a linear combination of normal distributions. We want to model the density of 3d point clouds using GMMs. This representation is used for point cloud processing tasks, such as registration or surface reconstruction. Our goal is to construct GMMs from point clouds specifically for applications in Deep Learning. In this thesis, different existing GMM construction algorithms are evaluated and new algorithms are designed to find the best algorithm for the given task. A gradient-descent-based method is currently in development. Additionally, a visualization tool is implemented aiding in reasoning about and inspection of three-dimensional GMMs.