Unsupervised Ego-Motion and Dense Depth Estimation with Monocular Video
Yufan Xu, Yan Wang, Lei Guo · 2018
In recent years, Deep Learning based method for 3-Dimension (3D) geometry perception tasks, like dense depth recovery, optical flow estimation and ego-motion estimation, is attracting significant attention. Inspired by recent advances in unsupervised strategies to learning from video datasets, we present a reasonable combination of constrains and a finer architecture, used for unsupervised ego-motion and depth estimation. Specifically, we introduce our effective neural networks Depth-Net (for monocular depth estimation) and Pose-Net (for ego-motion estimation), which are trained with monocular images. Depth-Net is proposed by us, improving the accuracy of estimation with as few parameters as possible. Finally, extensive experiments are implement on the KITTI driving dataset, proving our method outperforms some state-of-the-art results in unsupervised even supervised method.