Monocular depth estimation (MDE) is one of the core challenges encountered in the field of computer vision. MDE is the process of taking a singular image, and from it, estimate the approximate depth of the scene. The MDE process is quite a difficult task, as traditional depth estimation is carried out by making use of multiple views of the same object from different angles. MDE has seen much progress in terms of accuracy, and this is largely due to the use of deep learning techniques.
This project set out to find the state-of-the-art technique that would produce the best depth map for a given image. This was done by comparing the depth maps produced by each technique with the baseline depth map for each corresponding image, using evaluation metrics. With this information, the techniques could be arranged according to their accuracy of producing the current depth information for a given image. Another important component of this project is the refinement of the depth map of the image for 3D printing.
At the final stage, an experiment was carried out in order to evaluate the validity of the obtained results. This entailed printing the baseline depth map of an image and the corresponding depth map produced by the best technique in 3D, using a 3D printer. These 3D prints were then presented to visually impaired persons, who were asked to try to determine the object found within each of the two prints. Lastly, they were asked to indicate which of the pair they thought was the better one.
Course: B.Sc. IT (Hons.) Artificial Intelligence
Supervisor: Mr Dylan Seychell