Self-calibration and neural network implementation of photometric stereo

Abstract

This paper describes a new approach to neural network implementation of photometric stereo for a rotational object with non-uniform reflectance factor Three input images are acquired under different conditions of illumination. One illumination direction is chosen to be aligned with the viewing direction. We require no separate calibration object to estimate the associated reflectance maps. Instead, self-calibration is done using controlled rotation of the target object itself. Self-calibration exploits both geometric and photometric constraints. A radial basis function (RBF) neural network is used for non-parametric functional approximation. The neural network training data are obtained from rotations of the target object. Further, the method makes it possible to determine whether or not a given boundary point lies on an occluding boundary. The approach is empirical without needing a distinct calibration object and without making any specific assumptions about the surface reflectance. Experiments on real data are described.