Deep Neural Network Based Shape Reconstruction for Application in Robotics

Three-dimensional shape recovery and reconstruction, for classification and other applications, is an important task in Robotics. Shape From Focus (SFF) is one of the passive optical technique for 3D shape recovery, from a set of differently focused 2D images. It uses focus information as a cue to estimate 3D shape in the scene. Conventionally, images are taken at multiple positions along the optical axis of the imaging device, and stored in the image stack. This is followed by application of focus measure operators (FM). In order to reconstruct 3D shape of the object, best focused positions are obtained, by maximizing the focus curves obtained by FM application. Multiple FMs have been proposed in the literature but most of them are computationally expensive, since they have to process huge amounts of data. In this paper, Deep Neural Networks (DNN) have been employed to measure the amount of focus in the image stack with high accuracy. The results are compared with commonly used FMs by employing RMSE, Correlation and Q index.