Self-organizing elastic networks for generating a 3D model for many images
Manabu Motegi, Yukio Kosugi · Systems and Computers in Japan · 1999
When a human recognizes the shape of a 3D object, he observes the object from various directions. He seems to modify gradually the model that he has in his mind and finally recognizes the complete shape of the object. This article considers multiple 2D images taken from side views of a 3D object and proposes a self-organizing neural net that generates the horizontal cross section of the object, as well as its 3D shape. The neural net generates the object shape as if a rubber membrane encircles the object. It is a kind of elastic net where the shape is gradually generated by competition and collaboration (called consensus) among multiple elements. No special knowledge about the object shape is required. The shape is generated based on information on the image taken from two directions, the properties of the object rotation, and the a priori knowledge that the surface of the object is everywhere smooth. As the first step toward shape generation for the actual image, several simple 3D pseudo-objects are generated on a computer, and the generation process for the horizontal cross section and the 3D shape is verified for the pseudo-objects. © 1999 Scripta Technica, Syst Comp Jpn, 30(12): 106–115, 1999