Supervised and Unsupervised Methods in Depth Estimation

Abstract

Monocular depth estimation from single images has gained increasing attention in recent years, considering that this technique is one of the most important techniques in autonomous driving. Since the contrast and parameters of the indoor images internally differ from outdoor. this work presented two methods for optimizing depth estimation using convolutional neural networks. In the first method, the indoor images were dealt by mask prediction using an encoder-decoder structure (DRN) and by proposing three separate networks as depth estimator (ResNet-50, DenseNet-161 and ResNet-152). In the second method, which depends on outdoor images, depth estimated by CNN with no ground truth depth maps by using image reconstruction technique, with left-right disparity consistency check and autoencoder architecture (Resnet-18 model). Both proposed methods showed good performance compared to the reference studies.