Graphics, Imaging, and Games Lab

Image abstraction traditionally eliminates texture, flattening gradients and removing small-scale details. However, abstracting while preserving irregular silhouettes and medium-scale details can produce a richer abstracted image. We propose a variant of geodesic image filtering which preserves the locally strongest edges, leading to preservation of both strong edges and weak edges depending on the surrounding context.

Our contribution is to introduce cumulative range geodesic filtering, where the distance in the image plane is lengthened proportional to the color distance. We apply the new filtering scheme to abstraction applications in images and video, and demonstrate that it has powerful structure-preserving capabilities, especially regarding preservation and indication of irregular details. The basic technique, where every pixel is equally abstracted, is further extended with explorations of variable mask size based on spatial location, salience, intensity, and location combined with intensity.