In this thesis we address the problem of specular surface deflectometry-based depth estimation. Classical approaches, such as structure from motion, fail on specular surfaces -- that's why deflectometry-based methods are used. Typically these methods will use a sinusoidal fringe projection pattern and through use of frequency analysis and triangulation calculate the depth of a particular point. The drawback of these methods is that they require accurately calibrated system and multiple photos of the surface. In the thesis we propose method DeflectoDepth that is based on convolutional neural networks. It only needs one photo in order to work, for which it is able to predict both the depth and the mask of projected pattern. For evaluation purposes we prepared a data set CarDepth, where we used car bodies as the observed surface. Method DeflectoDepth achieves mean absolute error of 15.40 mm at depth estimation and precision of 98.5$\%$ and recall of 97.3$\%$ at mask prediction. We also developed a variant of the base method DeflectoDepth, which only predicts depth. This variant achieved mean absolute error of 13.44 mm.
|
