Learning probabilistic structure to group image edges for object extraction

We investigate exploiting the class specific information in the conventional perceptual edge grouping for the task of object extraction, since the domain information is usually available in practice. Instead of applying the classical Gestalt principles, we turn to learn a class specific probabilistic structure model from training images. During the learning, both geometrical and photometric features such as color and texture are fused. Experiments show the model is fairly robust to the intra-class variations of object as well as background clutters. Moreover, we design a novel saliency measure for the grouping based on the probabilistic structure model. The object extraction is formulated as an optimization problem which can be efficiently solved by the recently developed ratio contour algorithm. The effectiveness of the proposed method is demonstrated by the experiments on real images.