An efficient inference in meanfield approximation by adaptive manifold filtering (Machine learning & data mining)

A new method for speeding up the approximate maximum posterior marginal (MPM) inference in meanfield approximation of a fully connected graph is introduced. Weight of graph edges is measured by mixture of Gaussian kernels. This fully connected graph is used for segmentation of image data. The bottleneck of the inference in meanfield approximation is where the similar bilateral filtering is needed for updating the marginal in the message passing step. To speed up the inference, the adaptive manifold high dimensional Gaussian filter is used. As its time complexity is 0(ND), it leads to accelerating the marginal update in the message passing step. Its time complexity is linear and relative to the dimension and number of graph nodes. To improve the accuracy of segmentation, instead of the bilateral filter, the non-local mean filter is used. The proposed inference method is more accurate and needs less computations when compared to other existing methods.