Learning Monocular Depth Estimation with Unsupervised Trinocular Assumptions

Abstract

Obtaining accurate depth measurements out of a single image represents a fascinating solution to 3D sensing. \nCNNs led to considerable improvements in this field, and \nrecent trends replaced the need for ground-truth labels with \ngeometry-guided image reconstruction signals enabling unsupervised training. Currently, for this purpose, state-ofthe-art techniques rely on images acquired with a binocular \nstereo rig to predict inverse depth (i.e., disparity) according to the aforementioned supervision principle. However, \nthese methods suffer from well-known problems near occlusions, left image border, etc inherited from the stereo setup. \nTherefore, in this paper, we tackle these issues by moving to \na trinocular domain for training. Assuming the central image as the reference, we train a CNN to infer disparity representations pairing such image with frames on its left and \nright side. This strategy allows obtaining depth maps not \naffected by typical stereo artifacts. Moreover, being trinocular datasets seldom available, we introduce a novel interleaved training procedure enabling to enforce the trinocular \nassumption outlined from current binocular datasets. Exhaustive experimental results on the KITTI dataset confirm \nthat our proposal outperforms state-of-the-art methods for \nunsupervised monocular depth estimation trained on binocular stereo pairs as well as any known methods relying on \nother cues.