Unsupervised Scale Network for Monocular Relative Depth and Visual Odometry

Abstract

With the rapid development of deep learning and computer vision, learning-based monocular depth estimation and visual odometry have achieved increasingly remarkable results. However, there are few studies on the scale ambiguity of monocular depth estimation and visual odometry in an unsupervised network framework. Therefore, this article is to solve this thorny problem. We propose a joint unsupervised network framework that can provide necessary information to each other between different tasks to meet the needs of multiple tasks. To address the issue of scale ambiguity for learning-based monocular depth estimation, we propose a novel ScaleNet, an unsupervised scale network that provides scale information for the relative depths predicted by monocular depth networks, thereby recovering the absolute depths. Meanwhile, we propose a pseudo ground-truth scale generator and constrain the scale network by scale loss. The experimental results show that our monocular depth estimation results are competitive, and the scale network can provide reliable scale information for monocular depth networks. To address the challenge of scale ambiguity in learning-based monocular visual odometry, we propose a solution based on scale and optical flow to obtain the absolute scale of translational vectors using our depth alignment method. The experimental results show that our monocular visual odometry achieves state-of-the-art performance. The extensive experiments on the KITTI dataset for different tasks demonstrate the effectiveness and generalization of our proposed method.