DeepStereoBrush

We introduce a novel interactive depth map creation approach for image sequences which uses depth-scribbles as input at user-defined keyframes. These scribbled depth values are then propagated within these keyframes and across the entire sequence using a 3-dimensional geodesic distance transform (3D-GDT). In order to further improve the depth estimation of the intermediate frames, we make use of a convolutional neural network (CNN) in an unconventional manner. Our process is based on online learning which allows us to specifically train a disposable network for each sequence individually using the user generated depth at keyframes along with corresponding RGB images as training pairs. Thus, we actually take advantage of one of the most common issues in deep learning: over-fitting. Furthermore, we integrated this approach into a professional interactive depth map creation application and compared our results against the state of the art in interactive depth map creation.