Monocular Depth Estimation with Traditional Stereo Matching Information

Abstract

Monocular depth estimation is to perform pixel-level depth estimation on single perspective image, and the methods based on deep learning have shown superior effect for the challenging task. However, the supervised depth estimation methods need costly ground truth labels, which require professional equipment. To solve this problem, we study the unsupervised monocular depth estimation algorithm combining traditional stereo knowledge. We use the unsupervised network based on the reconstruction method as the baseline, which can be trained without direct supervision from ground truth depths leveraging image synthesis on sequences or multi-views images, so that it can improve the accuracy of prediction results without applying any annotations such as LIDAR depth in KITTI. Experiments results on the publicly database KITTI (Eigen split) have shown the promising effectiveness of the proposed method, and compared with state of the art unsupervised algorithms, it has achieved competitive performance in objective evaluation indicators.