Multi-View Reconstruction and Camera Recovery
using a Real or Virtual Reference Plane
Please download the ICCV '03 paper or my thesis for details

Please explorer movies and VRML models which I reconstructed with my method.
Your browser must support java 1.1.

Reconstructing a 3-dimensional scene from a set of 2-dimensional images is a fundamental problem in computer vision. A system capable of performing this task can be used in many applications in robotics, architecture, biometrics, human computer interaction and the movie and entertainment industry.

Most existing reconstruction approaches exploit one source of information to tackle the problem. This is the motion of the camera, the 2D images are taken from different viewpoints. We exploit an additional information source, a reference plane, which makes it possible to reconstruct difficult scenes where other methods fail. A real scene plane may serve as the reference plane. Furthermore, there are many alternative scenarios to obtain a (virtual) reference plane, like orthogonal directions in the scene or images taken with a parallel projection camera.

The main part of my work is about a novel multi-view reconstruction approach using a reference plane. The technique is applicable to three different feature types, points, lines and planes. The novelty of our approach is that all cameras and all features (off the reference plane) are reconstructed simultaneously from a single linear system of image measurements.