Depth Acquisition from Digital Images

Abstract

Introduction: Depth acquisition from digital images captured with a conventional camera, by analysing focus/defocus cues which are related to depth via an optical model of the camera, is a popular approach to depth-mapping a 3D scene. The majority of methods analyse the neighbourhood of a point in an image to infer its depth, which has disadvantages. A more elegant, but more difficult, solution is to evaluate only the single pixel displaying a point in order to infer its depth. This thesis investigates if a per-pixel method can be implemented without compromising accuracy and generality compared to window-based methods, whilst minimising the number of input images.

Method: A geometric optical model of the camera was used to predict the relationship between focus/defocus and intensity at a pixel. Using input images with different focus settings, the relationship was used to identify the focal plane depth (i.e. focus setting) where a point is in best focus, from which the depth of the point can be resolved if camera parameters are known. Two metrics were implemented, one to identify the best focus setting for a point from the discrete input set, and one to fit a model to the input data to estimate the depth of perfect focus of the point on a continuous scale.

Results: The method gave generally accurate results for a simple synthetic test scene, with a relatively low number of input images compared to similar methods. When tested on a more complex scene, the method achieved its objectives of separating complex objects from the background by depth, and produced a similar resolution of a complex 3D surface as a similar method which used significantly more input data.

Conclusions: The method demonstrates that it is possible to resolve depth on a per-pixel basis without compromising accuracy and generality, and using a similar amount of input data, compared to more traditional window-based methods. In practice, the presented method offers a convenient new option for depth-based image processing applications, as the depth-map is per-pixel, but the process of capturing and preparing images for the method is not too practically cumbersome and could be easily automated unlike other per-pixel methods reviewed. However, the method still suffers from the general limitations of the depth acquisition approach using images from a conventional camera, which limits its use as a general depth acquisition solution beyond specifically depth-based image processing applications.