Unsupervised Learning of Depth and Ego-Motion from Continuous Monocular Images

Unsupervised Learning of Depth and Ego-Motion from Continuous Monocular Images

In this study, the task of estimating depth is explored and also estimates continuous monocular images and optimizing and comparing two uncontrolled neural network structures namely DispNet and DispResNet, to determine a network structure that is more optimal. Photometric loss, minimal photometric loss, mask loss and smoothness loss are all components of loss functions for training depth and pose estimation neural networks. For the computation of photometric loss error caused through object motion and object occlusion on continuous images, a minimum photometric loss calculation method is proposed: the minimum value of photometric loss for each pixel point is taken, and then the mean value is computed as the minimum photometric loss, which minimizes the calculation error caused by occlusion, as well as other factors. The KITTI dataset assessment demonstrate that: the whole seven assessment parameters of depth estimation attain optimum value. Moreover, we show that our ego-motion network is able to predict camera tracks on long sequences of videos more closely than other algorithms.