Object cut as minimum ratio cycle in a superpixel boundary graph

A category-specific object cut method is proposed in this paper that utilizes both minimum ratio cycle optimization and superpixel segmentation. This method can find a non-self-intersecting cycle in the image plane which aligns well with the outer boundary of an object instance. Most existing approaches under the minimum ratio cycle optimization framework are used for unsupervised image segmentation. Directly applying their approaches will cause orientation ambiguity which makes the globally minimal solution unachievable. It is demonstrated that a modification on top-down classification information can alleviate this difficulty even it does not hold for traditional linear-energy object cut methods. PASCAL VOC 2007 segmentation dataset is used for experimental evaluation and improved performance is obtained when our method is compared with other competitive object cut algorithms.