A maximum likelihood approach to feature segmentation

摘要

Numerous processes in computer vision are based on the selection of features of interest (e.g. lines) which are then used to infer other information (e.g. vanishing point coordinates). The selection of features of interest is a difficult problem, first because of inaccuracy of the feature parameters, and second because of noise caused by the proximity of other, unrelated, features. A segmentation method based on the likelihood principle is described in this paper. The reliability of the process is optimized by using probabilistic models of the parameter uncertainty and of the noise created by the presence of the other features. It is compared with two popular tests; the first based on the Euclidean distance (called the neighbourhood test), and the second based on the Mahalanobis distance. Then, an example of this method for classifying lines with a vanishing point is described and tested on images from indoor scenes. This method has also been successfully applied to other vision segmentation problems (Brillault, A probabilistic approach to 3D interpretation of monocular images, Doctoral dissertation, City University, March (1992)).