Controlling growing snakes by using key-points

Abstract

In the present paper we address the problem of establishing contours in static monocular scenes. We propose to use features from unified edge and key-point detection to drive an active contour model. We present a framework for image segmentation consisting of: 1) detection of 1D and 2D intensity discontinuity, i.e. general edges and key-points; 2) reformulation of the active contour model to an expanding version, called a "growing snake"; 3) automatic initialization; 4) image and internal geometry driven expansion; and 5) controlled termination. The objective is to explore the potential of a fully automatic segmentation scheme. The performance and present limitations of growing snakes are demonstrated on images representative for diverse classes of scenes.