Depth Learner: An Advanced Learning Method with Edge Map for Monodepth2

When creating a deep learning model for estimating the depth of images, constructing a training dataset using stereo images presents a significant challenge. Therefore, using monocular images for depth estimation provides numerous benefits in terms of dataset acquisition. Monodepth2 is one of the prominent techniques for monocular depth estimation. By employing a self-supervised approach, Monodepth2 eliminates the need for ground truth, making the acquisition of the training dataset much easier. Nonetheless, a challenge faced by Monodepth2 is the issue of blurred boundaries in the output depth maps. To address this concern, the paper proposes a modified architecture of Monodepth2, resulting in enhanced accuracy and sharper boundaries in the output depth maps.