Roland Wilson, Multiresolution Gaussian Mixture Models: Theory and Application (February 28, 2000).
This paper introduces a new generalisation of scale-space and multiresolution pyramids, which combines statistical modelling with a spatial representation. The representation uses the familiar concept of multiple resolutions, but applied to a Gaussian mixture representation of the image; hence the title Multiresolution Gaussian Mixture Model (MGMM). It is shown that MGMM can approximate any probability density and can deal with the smooth motions that typically occur in image analysis and vision. An efficient recursive algorithm for computing MGMM representations, based on Monte Carlo Markov Chain methods, is presented. After a brief presentation of the theory, examples are used to show how MGMM can be applied to vision problems such as segmentation, stereopsis and motion analysis.