Unsupervised light field disparity estimation using confidence weight and occlusion-aware

Abstract

Light field disparity estimation is a crucial topic in computer vision. Currently, deep learning methods have shown significantly improved performance compared to traditional methods, especially supervised learning approaches. However, the high cost of obtaining real-world depth/disparity data for training greatly limits the generalization ability of supervised learning methods. In this paper, we propose an unsupervised learning method for light field depth estimation by utilizing confidence weights to evaluate the reliability of disparity features. First, during the disparity estimation and inference process, we introduce confidence weights to assess the reliability of disparity features, assigning higher weights to non-occluded and low-noise areas to effectively handle errors caused by occlusion and noise. Second, we design an occlusion-aware network to predict occluded regions in the views, which addresses the interference of occluded regions when computing unsupervised loss during training, thus enhancing the overall estimation accuracy. Extensive experimental results show that our method outperforms traditional methods and some of the latest unsupervised learning methods.