A characterizable shape-from-texture algorithm using the spectrogram

Abstract

Perspective-induced deformations on otherwise uniformly textured surfaces can be used to compute surface normals of objects from monocular images. This is shape-from-texture. Traditional shape-from-texture algorithms are based on image features like blobs and lines, and it is hard to predict how well the algorithms will work on real data. Newer algorithms are based on local spatial frequency representations, which can be characterized mathematically from beginning to end. We summarize our spectrogram-based algorithm, and show how we can characterize the performance of the algorithm based on the program parameters and the underlying texture.