A Synopsis of Monocular Depth Estimation

Abstract

AbstractDepth information is necessary for automated devices or software to accomplish tasks which require the knowledge of the surrounding environment without error. Monocular depth estimation uses a single 2D image to estimate depth, which makes it an exigent method. However, recent studies in monocular depth estimation based on convolutional neural networks indicate favorable performance in accuracy. Research on monocular depth training traverses through increasingly complex model architectures, and optimization of loss functions, all of these have recently helped to close the gap between the accuracy percentage of the traditional methods and trained models. In this review paper, we survey numerous existing literature on monocular depth estimation along with various datasets, as well as several supervised, unsupervised and semi-supervised algorithms. Furthermore, we figured out the drawbacks of the existing traditional methods and discussed contemporary methods using convolutional neural networks which may already have improved upon the said drawbacks. Lastly, we present our own findings on further scope in monocular depth estimation.KeywordsMonocular depthComputer visionDepth mapsSIDECNNUNet