Nils Einecke, Sven Rebhan, Julian Eggert, "Depth from Perspective Transformations", 5th HRI Global Workshop, 2008.

All binocular depth estimation algorithms need to apply some kind of matching process for correspondence search. Unfortunately, this search process is difficult because of ambiguities. One possibility to reduce ambiguities is to use 3-D models of the scene geometry. In this paper we interpret the scene as a composition of basic parameterizable surfaces. We present a general way of deriving formulas for perspectively mapping such surfaces from one stereo camera image to the second one. For estimating the model parameters we perform a search in the parameter space, which is guided by the error between the mapped and the original view. By means of the found model parameters depth values can be extracted. For searching the model parameter space we use the Hooke-Jeeves optimization which does not need an explicit gradient formulation and hence constitutes an easy way of circumventing the complex gradient formulas.