Occlusion-Aware Unsupervised Light Field Depth Estimation Based on Multi-Scale GANs

The estimation of depth from 4D light field images is a fundamental problem for perceiving and reconstructing environmental scenes. While learning-based methods have achieved remarkable results in this field, most of them rely on supervised learning, which faces significant challenges in real-world applications due to the lack of sufficient available ground truth depth maps. In this paper, we propose an unsupervised learning architecture based on a generative adversarial learning model for light field image depth estimation (OALFGAN). Specifically, our approach involves a multi-scale deep convolutional generative adversarial network learning system that includes a sparse-to-dense cascaded multi-scale generator and a discriminator, which decomposes the problem of generating high-quality images into more manageable sub-problems. To address the issue of violations of photometric consistency that may be caused by occlusion, we introduce a spatial-angular attention module that adaptively extracts view features with fewer occlusions and richer textures to generate more accurate disparity maps. Furthermore, we design a loss function that incorporates adaptive angular entropy consistency, symmetry loss, and edge-aware loss based on the distribution regularity and self-constraint of light field images to further optimize occlusion and disparity discontinuity issues and improve the reliability of the final depth prediction. Our proposed method demonstrates superior performance over existing methods on synthetic datasets, both quantitatively and qualitatively. Moreover, our proposed method exhibits excellent generalization performance on real-world datasets, demonstrating the effectiveness of our approach.