Bayesian object extraction from uncalibrated image pairs

This paper proposes a region-based algorithm for the extraction of foreground objects in uncalibrated images or video sequences. At the initialization step, a pair of images are overly segmented based on color and texture. Then, several dominant transforms that represent the relationship between two images are found, and each pixel is transformed by these transformations. By considering transform parameters and using the proposed area-ratio test, each pixel is assigned to a background or an object. We consider the resulting binary image as a binary random field, from which the likelihood of being foreground is computed for each initial segment. Using the likelihood model for each segment and the a priori assumption on the smoothness of labels, Bayesian approach is applied to label the segments. Experiments on various images and videos show promising results that the objects are clearly extracted.