LFDeNet: Light Field Depth Estimation Network Based on Hybrid Data Representation

This paper investigates an efficient CNN-based depth estimation for light field (LF) image in a hybrid data representation, named “LFDeNet.” We pay attention to the high dimensionality of LF data in the context of designing an efficient CNN structure for LF depth estimation task. Unlike most of existing CNN-based methods being applied on 2D subspace, i.e. 2D representation, we propose a hybrid CNN structure that works on different data representations to fully handle the 4D structure. Specifically, our network contains three novel feature extractor modules, allowing one to simultaneously exploit information from three types of 4D LF representation: spatial, angular, and epipolar image information. The experiments in this paper show that the proposed LFDeNet achieves competitive performance on the benchmark datasets and shows robustness with the occlusion and texture-less areas.