Magnetic contour tracing

Abstract

We present an interactive tool for image object boundary specification for the segmentation of unknown images. It feels like a freehand drawing tool, but behaves according to constraints related to the semantics of image object boundary formation. We find the path as the user traces by interpreting the user data as an approximation to the object boundary. From this, we derive the set of 8-connected paths which may be part of the boundary. To select the best path, we design a cost function in terms of the user data and the image data (e.g. gradient magnitude), and select a minimum cost path using a dynamic programming algorithm. A smooth path is produced that follows the user in low contrast regions and the object boundary otherwise. A method to tune the tool for specific conditions is described. We present quantitative results obtained for a simulated user using random data and qualitative results for a real user tracing in real images.<>