Accurate and Discernible Photocollages

Abstract

We propose a system for arranging images from a database into a collage that resembles some target image. These collages exploit large scale visual correspondences between the target image and the images in the database. We ensure that images of multiple sizes are used and are combined so that boundaries between images are not immediately apparently; as a result, the final collage consists of irregularly shaped image sections. The final collages contain a dynamic mixture of textures, images, and shapes that is in contrast to the geometric and regular character of many photomosaic techniques. In service of these tasks, we propose a fast scale-based method for querying an image library, a novel method for composing multiple images using geodesic distance Voronoi tesselations, and a novel base/detail method for shifting the colors of the final collage so that the target image is more accurately represented.