... reflection images. A convolutional neural network is trained in a supervised manner to learn the correspondence between RGB and depth images. Using the physical model obtained in the first step, an RGB image is reconstructed from the predicted depth map. To ensure the physical plausibility of the model’s...