Unpaired Learning of Dense Visual Depth Estimators for Urban Environments

Abstract

This paper addresses the classical problem of learning-based monocular depth estimation in urban environments, in which a model is trained to directly map a single input image to its corresponding depth values. All currently available techniques treat monocular depth estimation as a regression problem, and thus require some sort of data pairing, either explicitly as input-output ground-truth pairs, using information from range sensors (i.e. laser), or as binocular stereo footage. We introduce a novel methodology that completely eliminates the need for data pairing, only requiring two unrelated datasets containing samples of input images and output depth values. A cycle-consistent generative adversarial network is used to learn a mapping between these two domains, based on a custom adversarial loss function specifically designed to improve performance on the task of monocular depth estimation, including local depth smoothness and boundary equilibrium. A wide range of experiments were conducted using a variety of well-known indoor and outdoor datasets, with depth estimates obtained from laser sensors, RGBD cameras and SLAM pointclouds. In all of them, the proposed CycleDepth framework reaches competitive results even under a more restricted training scenario.