3D shape recovery by superquadrics model using object Silhouettes and stereo disparity

This paper describes a 3D shape reconstruction method using vision sensors targeted at domestic robotics applications. We propose a new method to fuse stereo disparity map and Shape from “Silhouette” (SFS). What we mean by silhouette in this paper is different from the existing silhouette definition. The silhouette here is not obtained from back projecting the object contour to the image plane but rather foreground/background stereo segmentation. Therefore, we impose the restraint that objects must be placed on a richly textured background. Furthermore, we use only three views to obtain the SFS. Preliminary results on domestic objects suggests that our method can distinguish objects such as cylinder, box and ball shapes.