Application of neural network in integration of shape from shading and stereo

In this paper, a simple and efficient approach is presented for the reconstruction of 3-D surfaces using the integration of shape from shading (SfS) and stereo. First, a new SfS algorithm is derived to obtain the depth-map of a 3-D surface using linear and generalized Lambertian reflectance model. Later, the accuracy of the depth-map is improved by integrating stereo depth data. The stereo sparse depth data are obtained at the points which have higher similarity score in the rectified pair of stereo images. A feed-forward neural network is used to integrate the SfS and stereo depth data due to its strong nonlinear function approximation property. The integration process is based on the correction of 3-D visible surface obtained from SfS using the stereo data. The experiments have been performed on real and synthetic images to demonstrate the usability and accuracy of the approach.