The Semi-Supervised Structure from Motion with Ground Plane Constraint

Abstract

Depth estimation of monocular views is an important task. We find that view synthesis methods often perform well in unsupervised deep recovery learning. Inspired by it, we propose a framework of depth recovery algorithm based on view synthesis. However, because view synthesis is based on the inherent constraints of the image pixels themselves, it is greatly affected by the “burst” conditions such as occlusion and illumination. In computer vision's domain, the algorithm of road segmentation and recognition from monocular images has been able to perform fairly well. Combining with the task of segmentation, we constrain depth estimation based on the premise that the spatial position of road pixels is on the ground. The experimental results show that with this rigid constraint in three-dimensional space, the task of depth estimation can perform better. The whole algorithm only needs the camera internal parameters to be known, and does not need road information, pixel depth information and other labeling, nor additional image information such as GPS, map, etc. The input raw data is used as the sequence of monocular data or video images, which greatly reduces the input of human resources and increases the level of automation of the system.