Binocular and Multiple View Stereo using Tensor Voting

Philippos Mordohai


Abstract

This talk will focus on an integrated approach to the derivation of scene description from stereo images, where both feature correspondence and surface reconstruction are addressed within the same framework.

The first part will be an overview of the binocular stereo approach developed by Mi-Suen Lee, while, in the second part, I will present my attempt to extend it to multiple views. In order to handle noise, occlusion and discontinuities, we adopt a tensor represenation for the data and tensor voting for information propagation. The key contributions of our approach, besides the intorduction of tensor voting in the field of stereo, are the way we impose constraints, the use of saliency instead of correlation as a criterion for the correctness of matches, and the execution of all processing in 3-D.


Maintained by Philippos Mordohai