Unsupervised Monocular Depth and Pose Estimation Using Multiple Masks Based on Photometric and Geometric Consistency

Abstract

With the rapid development of CNN-based deep learning methods, unsupervised monocular depth estimation and camera ego-motion estimation in consecutive frames have attracted much attention in recent years. In the process of photometric consistency-based image reconstruction, the pixel mismatching problem brings serious interference to model training. Compared with correctly matched pixels, mismatched pixels lead to severe influence on training loss. In this paper, to tackle the pixel mismatching problem, two novel binary photometric and geometric consistency-based auto-masks are presented to exclude abnormal errors in training loss. More significantly, the small loss in the forward and backward reconstructions is shoes to further reduce the interference of abnormal pixels. The combined evaluation results show that the proposed depth estimator achieves the state-of-the-art performance on the KITTI benchmark, and the results of visual odometry accuracy is competitive with the models that using optical flow or loop detection in traditional methods.