Self-Supervised Correlational Monocular Depth Estimation using ResVGG Network

Abstract

Self-supervised monocular depth estimation (SMDE) has recently received significant attention in computer vision. Leveraging the development of deep learning approaches, SMDE provides a solution to the applications of automation, navigation, and scene understanding. In this paper, we propose a novel training objective and learning network to perform a single image depth estimation in our convolutional neural network without the ground truth depth data. The proposed training objective enables the learning network to learn the stereo image correlation in training and estimates the image depth from a single input image in prediction. The proposed learning network ResVGG is a hybrid structure of Resnet50 and VGG-16. The proposed ResVGG has a similar performance as Resnet50 but needs much less computational costs. We demonstrate that our proposed method has competitive accuracy comparing to the current state-of-the-art on KITTI dataset and achieves the frame rates of 33 frame per second (FPS) in prediction using a single NVIDIA GTX 1080 GPU. Furthermore, the proposed method can potentially support visual odometry depth estimation.