Object segmentation is a challenging task in computer vision.Without utilizing any high-level knowledge about expected objects, image segmentation based on low-level information such as edges or region statistics, often performs poorly in the presence of noises, background clutter, object occlusions, etc.In numerous studies,the addition of shape prior information to the segmentation process has shown to significantly improve segmentation results.However, most of these methods are recognition-based segmentation.They only segment objects of a known class in the image according to their possible similar prior shapes.In this work, we consider more general case where we do not know the class of object in the test image.Suppose that prior knowledge given by a shape category is associated with a set of different object classes, and focus on the problem of how to exploit such shape priors to guide object segmentation and recognition.