PDE-Net: Pyramid Depth Estimation Network for Light Fields

Abstract

Depth information is critical for highly developed robotic systems which need the global perception of their surroundings. A light field camera records with a large amount of data, increasing potentials for the depth estimation improvements. This paper addresses the challenge of performing depth estimation from a light field image: preserving depth discontinuity, which provides robotics with the more precise perception. We design an end-to-end trainable Pyramid Depth Estimation Network PDE-Net, by introducing the multi-scale feature merging from stacked images along four directions of light fields to perform the coarse-to-fine estimation. Extensive experiments on both synthetic light field datasets and real-world light field datasets demonstrate the superior performance of the proposed learning-based method compared to the state-of-the-art methods.