New shape from Shading methods

Abstract

Shape from Shading is perhaps the most difficult topic to deal with in Artificial Vision: several researchers have faced it using different approaches. The most part of these methods are based on the Horn algorithm so they require very heavy regularity assumptions about the perceived objects' shape and are computationally expensive.

The use of neural networks may be a good solution to the drawbacks of the classical approaches to SFS: in fact a neural network is able to solve estimation problems through a process of learning from a few meaningful examples requiring a very low computational cost.

Two different neural approaches are proposed by the authors: the first one consists of a cascaded architecture made up by a first stage named BWE (Boundary Webs Extractor) which is aimed to extract a brightness gradient map from the image, followed by a backpropagation network that estimates the geometric parameters of the object parts present in the acquired scene. The second approach is based on the extraction of the boundary webs map from the image and its comparison with boundary webs maps exhaustively generated from synthetic superquadrics. A purposely defined error figure has been used to find the best match between the two kinds of maps.

A functional comparison between the two systems is described and the quite satisfactory experimental results are presented.