A description of a model for accumulating segmentation information over multiple images is presented. The model combines geometric information with segmentation information to improve predicted appearance of objects. An extensive example using polyhedral object representations is described. Statistical results are shown over a set of 40 real images, verifying the underlying assumptions of the segmentation model concerning spatially distributed segmentation information over object edges. Evidence is accumulated in support of the hypothesis that segmentation varies significantly because of the interference between scene objects as well as viewpoint and illumination.