https://sites.google.com/site/hgjeoncv/home/depthfromlf_cvpr15This paper introduces an algorithm to estimate accurate depth map from a lenslet
light field camera. Our algorithm estimates multi-view stereo correspondences at
sub-pixel accuracy using a cost volume.
Our key idea to build accurate costs is threefold.
First, sub-aperture images are exactly displaced using the phase shift theorem.
Second, gradient costs are adaptively aggregated using the angular coordinate of the light field.
Third, feature correspondences between the sub-aperture images are utilized as an additional constraint.
With the cost volume, a multi-label optimization propagates and corrects depth map at weak texture regions.
Finally, we iteratively refine local depth map by fitting local quadratic
function to estimate a non-discrete depth map. Since a micro-lens image contains
unexpected distortions, we also present a method to correct the error.
The effectiveness of our algorithm is demonstrated
through challenging real world examples, with comparisons to the performance
of state-of-the-art depth estimation algorithms for light field images.