Depth from Monocular Images using a Semi-Parallel Deep Neural Network (SPDNN) Hybrid Architecture

Computing pixel depth values provide a basis for understanding the 3D geometrical structure of an image. As it has been presented in recent research, using stereo images provides an accurate depth due to the advantage of having local correspondences; however, the processing time of these methods are still an open issue. To solve this problem, it has been suggested to use single images to compute the depth values but extracting depth from monocular images requires extracting a large number of cues from the global and local information in the image. This challenge has been studied for a decade and it is still an open problem. Recently the idea of using neural networks to solve this problem has attracted attention. In this paper, we tackle this challenge by employing a Deep Neural Network (DNN) equipped with semantic pixel-wise segmentation utilizing our recently published disparity post-processing method. Four models are trained in this study and they have been evaluated at 2 stages on KITTI dataset. The ground truth images in the first part of the experiment come from the benchmark and for the second part, the ground truth images are considered to be the disparity results from applying a state-of-art stereo matching method. The results of this evaluation demonstrate that using postprocessing techniques to refine the target of the network increases the accuracy of depth estimation on individual mono images. The second evaluation shows that using segmentation data as the input can improve the depth estimation results to a point where performance is comparable with stereo depth matching.