Multi-View Image Compositions

Abstract

The geometry of single-viewpoint panoramas is well understood: multiple pictures taken from the same viewpoint may be stitched together into a consistent panorama mosaic. By contrast, when the point of view changes or when the scene changes (e.g., due to objects moving) no consistent mosaic may be obtained, unless the structure of the scene is very special. Artists have explored this problem and demonstrated that geometrical consistency is not the only criterion for success: incorporating multiple view points in space and time into the same panorama may produce compelling and informative pictures. We explore this avenue and suggest an approach to automating the construction of mosaics from images taken from multiple view points into a single panorama. Rather than looking at 3D scene consistency we look at image consistency. Our approach is based on optimizing a cost function that keeps into account image-to-image consistency which is measured on point-features and along picture boundaries. The optimization explicitly considers occlusion between pictures. We illustrate our ideas with a number of experiments on collections of images of objects and outdoor scenes.

Files

Additional details