Unsupervised learning of light field depth estimation with spatial and angular consistencies

Learning-based depth estimation from light fields has made significant advances in recent years, however, most of these work abandon the traditional non-learning based formulations and start over with an end-to-end deep network framework. Actually traditional methods have also presented many effective and evident depth cues or constraints for light field depth estimation. In this paper, we try to combine the advantages of the learning and non-learning strategies, and incorporate the traditional light field constraints into an unsupervised framework by using a learnable approximation scheme to yield differentiable unsupervised loss functions. Specifically, we first propose an unsupervised coarse-to-fine network architecture for light field depth estimation, and then design an adaptive spatio-angular consistency loss combined with the differentiable versions of modified traditional constraints. Comparative experiments demonstrate that our method is superior to the state-of-the-art unsupervised methods.