Single image depth estimation: An overview

We review solutions to the problem of depth estimation, arguably most important subtask in scene understanding. focus on single image estimation problem. Due its properties, is currently best tackled with machine learning methods, successfully convolutional neural networks. provide an overview field by examining key works. examine non-deep approaches that mostly predate deep and utilize hand-crafted features assumptions, more recent works use techniques. The a supervised fashion absolute or relative information acquired from human sensor-labeled data, unsupervised way using unlabeled stereo images video datasets. also study multitask combine related tasks such as semantic segmentation surface normal estimation. Finally, we discuss investigations into mechanisms, principles, failure cases contemporary solutions.