Superquadric-based scene description

This paper is concerned with describing scenes consisting of piles of simple but unknown objects from dense range image. What differentiates this work from other approaches is that it is purely data driven, using concepts such as symmetry, viewpoint independence, and solidity to hypothesize the unknown shape and size of the objects. We use the superquadric model for scene description so that the method is applicable for general 3-D objects. The unknown shapes and sizes of the objects in the scene are verified by an algorithm which uses the enclosing boundings and the central moments for initial estimate of sizes and poses, and then the refined estimations are obtained by a nonlinear least squares technique. The final scene description is based on a relational graph of primitives. This techniques may be applied in situations where dimensioned CAD models are unavailable.<>