Probabilistic point set matching with Gaussian mixture model

In this work, we propose a variational approximation approach in combination with isotropic Gaussian mixtures with regard to each individual transformed points for point set matching problems. A variational inference algorithm is formulated to update the posteriors of the random variables in sequence until a local optimum is reached. The probabilistic framework explicitly accounts for matching uncertainty and is thus less prone to local optima. Furthermore, the Gaussian mixtures with anisotropic covariance are also proposed for the modeling of spurious points instead of the one uniform distribution. The experimental results show that the combination of variational approximation with mixture model provides our algorithm with comparable performance of accuracy and robustness to other registration algorithms in the presence of outliers.