Depth from relative normal flows

Abstract

Most of the depth from image flow algorithms has to rely on either good initial guesses, or some assumptions about the object surfaces to achieve solutions that agree with the physical world. Waxman and Sinha point out that those restrictions can be relaxed if depth is computed from a relative image flow field. Since image flow determination is relatively much more difficult than normal flow determination, it is of interest to develop an algorithm to recover depth from normal flows. In this paper, we have shown that similar results can be obtained from relative normal flow fields as from relative image flow fields. We have implemented a normal flow estimation algorithm, and applied our algorithm to recover depth from intensity images.