Investigating the accuracy of shape recovery from light field cameras

Abstract

The development of light field sensor technology has the potential to bring three-dimensional imaging into our everyday life. This breakthrough is limited by the lower resolution of the depth component in 3D images in comparison to the mega-pixel resolution of 2D images. In current systems the depth component of 3D reconstructions from light field cameras is noisy and only accurate at specific locations, such as color boundaries and highly-textured regions. In this thesis we aimed to better understand the problem by analyzing the accuracy of shape recovery with light field cameras. We implemented the plane sweep stereo algorithm in MATLAB to reconstruct 3D images with different numbers of views and noise conditions. As the number of views increases, correlation is performed amongst more views, but the finite number of pixels available causes the resolution of each view to decrease. Based on our experiments we determined that there is a trade off between these two effects, resulting in an optimal number of views. Moreover, the optimal number of views increases as more noise is added to the images.