Automatic relevance determination for the estimation of relevant features for object recognition

Object recognition from 2D images is a highly interesting problem. The final goal is to have a system which can recognize thousands of different categories like human beings do. However, hand labelling the 2D training images in order to segment the foreground (object) from the background is a very tedious job. Because of this reason, in recent years, intelligent systems which can learn object categories from unlabelled image sets have been introduced. In this case, an image is labelled by the objects which are present in the image but the objects are not segmented in the image. The main problem in this case is that the object and the background are used altogether in such unsupervised systems and segmentation must be performed by the system itself. Automatic Relevance Determination (ARD) [8] is a method which will be investigated in this study in order to segment foreground and background in an unsupervised object category learning system.