Direct Fitting of Gaussian Mixture Models

Authors: Keselman

Year: 2019

Notes:

Usually model fitting with mesh is done by sampling points
Propose to apprixomate the surface with a Gaussian Mixture Model (GMM)
A GMM is a sum of a gaussian with normalization factors such that $\sum \lambda_i = 1$ $$ g(x)=\sum_{i=1}^K \lambda_i \mathcal{N}\left(x ; \mu_i, \Sigma_i\right) $$
How to get covariance of triangle? cf Principal Component Analysis in CGAL : use the vertices to compute a distribution centered on the center of mass
Able to perform both P2D (point to distribution) and D2D registration
Better than ICP on the stanford bunny (less outliers)
Perform VO on the TUM dataset using D2D registration (more in On-Manifold GMM Registration)