Automating joiners

Pictures taken from different view points cannot be stitched into a geometrically consistent mosaic, unless the structure of the scene is very special. However, geometrical consistency is not the only criterion for success: incorporating multiple view points into the same picture may produce compelling and informative representations. A multi viewpoint form of visual expression that has recently become highly popular is that of joiners (a term coined by artist David Hockney). Joiners are compositions where photographs are layered on a 2D canvas, with some photographs occluding others and boundaries fully visible. Composing joiners is currently a tedious manual process, especially when a great number of photographs is involved. We are thus interested in automating their construction. Our approach is based on optimizing a cost function encouraging image-to-image consistency which is measured on point-features and along picture boundaries. The optimization looks for consistency in the 2D composition rather than 3D geometrical scene consistency and explicitly considers occlusion between pictures. We illustrate our ideas with a number of experiments on collections of images of objects, people, and outdoor scenes.