Abstract

Light field disparity estimation is an important task in light field applications. However, how to efficiently utilize the high-dimensional light field data is a very worthy issue to investigate. Besides, existing supervised deep learning based algorithms are limited to scenes with ground truth disparity for training. In this paper, we propose a light field disparity estimation network which adopts a cascade cost volume architecture and can predict disparity maps in a coarse to fine manner by fully exploring the geometry characteristics of sub-aperture images. In addition, we design a combined unsupervised loss to train our network without a ground truth disparity map. Our combined loss consists of occlusion-aware photometric loss and edge-aware smoothness loss which can bring targeted performance improvements in occlusion and textureless regions, respectively. Extensive experiments demonstrate that our approach can achieve better results compared to existing unsupervised disparity estimation method and show better generalizability compared to supervised methods.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call