Depth Estimation from a Single Image using Multi Stream and Scale Deep Learning

Abstract

In this paper, we propose a method to estimate a depth map from a single image by combining the strengths of multi-scale and multi-stream deep neural networks. In the first scale, we use an encoder to allow the system to learn two types of inputs: an RGB image and image patches based on their superpixel. The second network (decoder) is accomplished to get a fine depth map by propagating the previous output with those corresponding RGB image. By combining an RGB image and superpixel patches, we can achieve a reliable feature depth map in the encoder network with a small number of training data. Consequently, we can enhance the ability of the refining network to predict the final depth. The effectiveness of our method is shown by real image experiments.