Stereovision

The different perspectives of our two eyes lead to slight relative displacements of objects (disparities) in the two monocular views of scene. The human visual system is able to

use the disparities for depth-estimation.
merge both monocular views into a cyclopean view of the scene.

The correct and fast estimation of disparities is a difficult problem, since various additional image variations occur between the left and right view. Differences might be caused by occlusions of objects, specular reflections, which move independently of the surfaces of objects, and various other image variations.

Stereovision has long been researched; common computational approaches include feature-based, area-based and phase-based methods. All these methods have their intrinsic problems, caused by the very assumptions inherent in these approaches.

Here, a new, simple and fast algorithm is described, which is based on the coherence-detection between simple disparity units. The neural network implementing the algorithm creates the cyclopean view of the scene as a simple by-product of the disparity calculations.

Further infos:

Try it out: Disparity calculations via the World Wide Web, with your stereo data!
Paper: "Stereovision by Coherence Detection" | Online-version | Download (522k) |
Some results obtained with coherence-based stereo

Stereovision and the cyclopean view
Comparision: Classical cooperative schemes and the new approach
Some VRML-1.0 files, calculated via the new approach

Comments are welcome! © 1997 by Rolf Henkel - all rights reserved.