Light fields, being inherently a 4D function cannot be mapped onto the 2D sensor in a single image without loosing out on resolution. A natural way to overcome this barrier is to capture multiple images to record the light field. However, this method only works for static scenes, therefore the resolution problem stays unresolved, it only gets transformed from the domain of low spatio-angular resolution to a problem of low temporal resolution.

In this work, we leverage the redundant nature of light fields to recover them at higher resolution by first capturing a set of well-chosen images, and later reconstructing the LF from these images using some prior-based algorithms. We achieve this in two ways. In the first method, we capture multiplexed light field frames using an electronically tunable programmable aperture and later recover the light field using a motion-aware dictionary learning and sparsity based reconstruction algorithm. The number of adjacent multiplexed frames to be used during the recovery of each light field frame is decided based on the applicability of the static scene assumption. This is determined using optical-flow and forms the basis of our motion-aware reconstruction algorithm. We also show how to optimize the programmable aperture patterns using the learned dictionary.

Our second method utilizes focus stacks to computationally recover light fields post-capture [1] . However our method differs from [1] in the following ways. (i) We obtain the entire focus-aperture (45 focus and 18 aperture settings) stack by capturing just a few (about 8−16) images and computationally reconstructing images corresponding to all other focus-aperture settings, while [1] capture the entire focus stack corresponding to a given aperture setting (ii) Since we recover the focus stack at smaller aperture settings as well, we can produce LFs at finer angular resolutions. We call our method 'Compressive Epsilon Photography' since we capture few (compressive) images with slightly varying parameters (Epsilon Photography) and post-capture computationally reconstruct images corresponding to all other missing parameter combinations. The recovered LF has spatial resolution corresponding to the sensor resolution of the camera and can recover any angular view which lies inside the aperture.

[1] A. Levin and F. Durand, "Linear view synthesis using a dimensionality gap light field prior," in IEEE conf. Computer Vision and Pattern Recognition, pp. 1831-1838, 2010