Multi-resolution Dense Residual Networks with High-Modularization for Monocular Depth Estimation

Abstract

Deep-learning neural networks (DNN) have been acknowledged to capably solve the ill-posed monocular depth estimation problem in self-drive applications. In this paper, we proposed a dense residual multi-resolution supervised DNN toward accurate monocular depth estimations for traffic landscape scenes. The proposed DNN is constructed by regularly integrating the dense residual short-cut connections into the multi-resolution backbone. However, since some implicitly influential features cannot be viable at the end of learning, the DNN structure for generating the details of estimated monocular depths shall not be too deepened. Basically, the structural depth of DNN can be suppressed by effectively exploiting the functional residual connections. In the proposed DNN structure, the amount of short-cut connections can be moderate through rational employments. Particularly, for achieving high modularization, we address three-layered modules to generate the adequate levels and layers in which the results can easily be controlled to meet a requested prediction/inference accuracy. Therefore, the visualization and quantitative results can demonstrate the superiority of the proposed DNN to other compared DNNs for street landscape experiments.