Shape model fitting using non-isotropic GMM

We present a Mean Shift algorithm for fitting shape models. This algorithm maximises a posterior density function where the likelihood is defined as the Euclidean distance between two Gaussian mixture density functions, one modelling the observations while the other corresponds to the shape model. We explore the role of the covariance matrix in the Gaussian kernel for encoding the shape of the model in the density function. Results show that using non-isotropic covariance matrices improve the efficiency of the algorithm and allow to reduce the number of kernels to use in the mixture without compromising the robustness of the algorithm.