Attention-based Self-supervised Learning Method for Monocular Depth Estimation

Abstract

Depth information is essential for the device to locate itself in 3-dimensional space. Convolutional Neural Network is used for monocular image depth estimation and the performance is significantly improved compared to traditional hand-craft methods. However, it takes a lot of time and physical costs to generate the large dataset with accurate Ground Truth required for learning. Therefore, methods of estimating depth based on self-supervised learning or unsupervised that do not require labeling have been proposed. In addition, deep neural networks are confronted by the accuracy degradation by stacking convolutional layers. For better performance of a deep neural network, this paper focused on the relationship between channels. In this paper, the method of estimating the monocular depth is proposed to solve these problems by using self-supervised learning that produces labels on its own. The enhanced encoder with squeeze and excitation layers is used for the better feature responses to estimate depth. The extra layers added in the residual bottleneck helped to improve the quality of representations presented by the network. The proposed self-supervised learning method produces labels with algorithms that predict the shape of the input image from the view of time before and after the input image sequence. In particular, automatic masking for removing pixels that violate the camera movement and the minimum value of reprojection error were used to allow more accurate depth estimation even when there is no motion in the video or objects that are obscured. KITTI dataset was used to evaluate the learning and performance of the network, and it was confirmed that high accuracy and sharp depth maps were printed compared to other methods of estimating monocular depth.