Pavel Kisilev (IBM Research Haifa)
Tuesday, 10.12.2013, 11:30
In this talk we present various research topics in medical imaging and related to them computer vision methods that are developed at IBM Haifa Research Lab. In the first part of the talk we describe various medical imaging problems and briefly hint on their solutions. The second part of the talk will be devoted to a newly developed general method for figure-ground segmentation. The method combines a bottom up approach of generating multiple highly plausible labelings of an image with a top-down model for reranking these segmentations, giving competitive results on a number of challenging datasets. Contrary to competing state-of-the-art methods for figure/ground segmentation that rely on complex models of foreground and boundary, our approach uses a simple binary pairwise graphical-model for bottom-up inference which relies on features computed on superpixels in the image and can be learned efficiently. The ranking stage, trained using features that rely on both the image evidence and inferred segmentation, discriminates between good and bad figure/ground labelings produced from the bottom-up inference. We will present examples of application of the proposed segmentation method to natural images and to difficult medical image cases.