Estimating Depth from Single Image Using Unsupervised Convolutional Network

To improve the accuracy of monocular depth estimation by deep learning, this paper proposes a method of unsupervised convolutional neural network for depth estimation from one single image. By introducing residual structure, dense connection structure and short-cut connection in the encode-decode network structure, the single image depth estimation convolutional neural network is improved and the learning efficiency and performance of the network are improved, and the convergence speed of the network is accelerated. Secondly, combined with the loss metrics such as gray similarity, disparity smoothing and left and right disparities matching, a more efficient loss function is designed, which effectively reduces the influence of image illumination factors, suppresses the discontinuity of image depth and ensures the consistency of left and right disparities. Thereby, the robustness of the depth estimation is improved. Finally, this method realizes the end-to-end depth estimation from single image, where stereo image sequences are used as the training data and the depth information in the single image scene can be estimated without the target depth supervision information. Experiments and comparisons on the KITTI and Cityscapes datasets with TensorFlow framework prove the effectiveness of the proposed method with higher accuracy and faster convergence. © 2020, Beijing China Science Journal Publishing Co. Ltd. All right reserved.