Researchers in the Visual Geometry Group of the University of Oxford’s (Department of Engineering Science) have developed reliable and flexible geometrical techniques for reconstructing Virtual Reality worlds from single and multiple perspective images.
An important feature of the work is that no knowledge of the camera position and its properties is required at any time: they can use uncalibrated images.
In particular, techniques based on extensive use of Projective Geometry have been developed and incorporated in computer algorithms capable of modelling perspective distortion and recovering the complete three-dimensional information from a single perspective image.
The reconstruction process is difficult because during projection onto a flat image many of the clues about the three dimensional nature of the viewed scene are lost. However, information lies encoded in the perspective distortion in the image, which makes objects farther away look smaller. In order to recover complete 3D structure, this distortion needs be modelled correctly, which in typical applications requires knowledge of the camera’s position and lens.
A delightful result of this ability to handle single, uncalibrated images is that it is possible to reconstruct scenes in paintings conforming to the rules of Linear Perspective discovered by Filippo Brunelleschi at the beginning of 15th century in Florence.
Several examples of such paintings can be found in the Italian Renaissance period. The computer is able to optimally process the geometric information contained in the painting, interpret the structure of the represented scene and construct interactive Virtual Reality models and animations.
More detail and some astonishing videos are available on the web site below.