Purdue Research Foundation

Zygmunt Pizlo

Yunfeng Li

Tadamasa Sawada

Robert M Steinman

Systems and methods for processing a pair of 2-D images are described. In one example, a stereoscopic set of images is converted into a collection of regions that represent individual 3-D objects in the pair of images. In one embodiment, the system recovers the 3-D point P for each point p that appears in both images. It estimates the 3-D orientation of the floor plane, and the image capture planes and their height from the floor. The system then identifies the collection B of points P that do not represent points on the floor and generates a projection C of B onto a plane parallel to the floor. It blurs the projection C and identifies peaks in the blurred image, then fits symmetric figures to the points in C around the identified peaks. The system projects the 3-D figures associated with the symmetric figures back onto the 2-D images.

Figure ground organization of 3-D scenes

Figure-ground organization of 3-D scenes

A system processes 2D images of 2D or 3D objects, creating a model of the object that is consistent with the image and as veridical as the perception of the 2D image by humans. Vertices of the object that are hidden in the image are recovered by using planarity and symmetry constraints. The 3D shape is recovered by maximizing 3D compactness of the recovered object and minimizing its surface area. In some embodiments, these two criteria are weighted by using the geometric mean.

Reconstruction of shapes of near symmetric and asymmetric objects

The present disclosure describes a system and method for transforming a two-dimensional image of an object into a three-dimensional representation, or model, that recreates the three-dimensional contour of the object. In one example, three pairs of symmetric points establish an initial relationship between the original image and a virtual image, then additional pairs of symmetric points in the original image are reconstructed. In each pair, a visible point and an occluded point are mapped into 3-space with a single free variable characterizing the mapping for all pairs. A value for the free variable is then selected to maximize compactness of the model, where compactness is defined as a function of the model's volume and its surface area. 'Noise' correction derives from enforcing symmetry and selecting best-fitting polyhedra for the model. Alternative embodiments extend this to additional polyhedra, add image segmentation, use perspective, and generalize to asymmetric polyhedra and non-polyhedral objects.

Growing community of inventors

Indianapolis, IN, United States of America

Zygmunt Pizlo