3D printing of 2D images

Abstract

Monocular Depth Estimation is often considered as an ill-posed problem within the world of computer vision. Extracting depth information from a singular 2D image could prove to be very useful in various situations, such as scene reconstruction, 3D object recognition, object detection and classification. Traditional techniques of carrying out depth estimation rely on the presence of multiple views of the same image from different angles. Over the past decade, progress has been made with regards to trying to be able to produce more accurate results, mostly with the help of deep learning techniques. Each implementation varies slightly in either the network architecture, or the training strategy implemented. The aim of this dissertation was to find depth estimation techniques, and evaluate them in order to find the most accurate depth estimation technique. This technique would then predict the depth maps of a set of images, which will be converted into a set of 3D prints by means of a 3D printer. These 3D prints will then be used in an experiment to see how effective depth estimation techniques actually are, and to see if such an implementation of printing images would be beneficial to visually impaired people. The results obtained throughout this dissertation were that the DPT-Hybrid technique was the best depth estimation technique, when compared to its competitors LapDepth and DenseDepth. In terms of the experiment, it was shown that making use of such depth estimation techniques is quite effective, as each 3D print was able to portray what the image originally was. On the contrary, some more progress has to be made, so that this implementation would be useful for visually impaired people, as currently it would require additional help from another individual in order to determine the objects found within the 3D print of the image.