Structure tensor-based depth estimation from light field images

This thesis presents a novel framework for depth estimation from light eld images based on the use of the structure tensor. A study of prior knowledge introduces general concepts of depth estimation from light eld images. This is followed by a study of the state-of-the art, including a discussion of several distinct depth estimation methods and an explanation of the structure tensor and how it has been used to acquire depth estimation from a light eld image. The framework developed improves on two limitations of traditional structure tensor derived depth maps. In traditional approaches, foreground objects present enlarged boundaries in the estimated disparity map. This is known as silhouette enlargement. The proposed method for silhouette enhancement uses edge detection algorithms on both the epipolar plane images and their corresponding structure tensor-based disparity estimation and analyses the di erence in the position of these di erent edges to establish a map of the erroneous regions. These regions can be inpainted with values from the correct region. Additionally, a method was developed to enhance edge information by linking edge segments. Structure tensor-based methods produce results with some noise. This increases the di culty of using the resulting depth maps to estimate the orientation of scenic surfaces, since the di erence between the disparity of adjacent pixels often does not correlate with the real orientation of the scenic structure. To address this limitation, a seed growing approach was adopted, detecting and tting image planes in a least squares sense, and using the estimated planes to calculate the depth for the corresponding planar region. The full framework provides signi cant improvements on previous structure tensorbased methods. When compared with other state-of-the-art methods, it proves competitive in both mean square error and mean angle error, with no single method proving superior in every metric