A Comparative Study of Models for Monocular Depth Estimation in 2D Images

Monocular depth estimation has been a challenging topic in the field on computer vision. There have been multiple approaches based on stereo and geometrical concepts to try and estimate depth of objects in a two-dimensional field such as that of a plain photograph. While stereo and lidar based approaches have their own merits, there is one issue that seems recurrent in them, the vanishing point problem. An improvised approach to solve this issue involves using deep neural networks to train a model to estimate depth. Even this solution has multiple approaches to it. The general supervised approach, an unsupervised approach (using autoencoders) and a semisupervised approach (using the concept of transfer learning). This paper presents a comparative account of the three different learning models and their performance evaluation.