Fast Light Field Disparity Estimation via a Parallel Filtered Cost Volume Approach

Abstract

The Guided Light Field Cost Volume (GLFCV) is a light field disparity estimation algorithm designed for (GPU) parallelization by refactoring the process, such that costly optimizations that combine and refine depth maps are simplified. The algorithm involves shearing the light field over a range of disparities and computing a cost volume for each sheared sub-aperture image. A guided filter is then run on the computed cost for each disparity. For the final disparity estimate, each pixel is assigned disparity associated with the lowest filtered cost. Execution time and accuracy were evalulated on Lytro Illum imagery and also on synthetic light fields from the HCI 4D light field benchmark. The approach presented executes in four seconds (on an NVIDIA Titan XP), with only 10% of pixels in the disparity estimate deviating from ground truth. This compares favourably to existing accurate approaches, whose execution time is in the order of minutes.