Tech Report CS-20-01

Fast and Accurate 4D Light Field Depth Estimation

Numair Khan, Min H. Kim and James Tompkin

August 2020

Abstract:

We present an algorithm for accurate depth estimation from 4D light fi elds that runs almost an order of magnitude faster than classical methods. Our p roposed approach use epipolar-plane image edges to estimate sub-pixel disparity at a small set of pixels in the 4D space. By optimizing constraints at these pix els we are able to diffuse the sparse set in an occlusion-aware manner to obtain dense disparity maps. Qualitative and quantitative results on both synthetic an d real-world light fields show that we have comparable, or better performance th an existing methods, while being significantly faster (8--11x) than current non- learning-based methods.

(complete text in pdf)