Superquadric is a parametric, volumetric primitive, that can be used to represent real world objects with only a few parameters. My dissertation work consists of representing multi-part complicated objects as well as real scenes with both regular and deformable sueprquadrics. In the case of real scenes, a multi-view representation scheme is constructed to reduce the effect of occlusions inherited in single-view range images, and improve the accuracy and confidence of superquadrics recovered from single-view images. For multi-part obejcts, 3D pre-constructed triangulated models are used to eliminate the visibility ambiguities contained in single-view images. A curvature based 3D part decomposition algorithm is proposed to decompose the multi-part objects into single, simpler objects. Sueprquadrics are then recovered from those single-part objects.

To represent multi-part objects with superquadircs, we propose the following strategies:

• Deformable superquadric recovery from unstructured 3D data.
• A multi-view superquadric representation framework.
• A 3D part decomposition algorithm based on curvature analysis.
• Superquadric representation of multi-part objects.

The proposed algorithms can be applied to various tasks in computer vision, reverse engineering, and computer graphics, including object recognition, 3D object modling, scene description, and virtual reality.