An occlusion light field sparse Bayesian learning model for view synthesis

Given a set of captured views with known positions, our goal is to obtain different views from new positions. However, synthesizing novel views from new positions is challenging since occlusion in real-world scenes is complex and ubiquitous. In this paper, we describe a method for synthesizing a novel view of an occluded scene, that is, an occlusion light field (OLF) sparse Bayesian learning network (OLiFi-Net). Specifically, we break down the process into OLF parameterization and interpolation reconstruction components. For the first component, we utilize a sparse Bayesian learning approach to establish an OLF expression. This expression can then be used to derive the convolution interpolation kernel function. For the second component, the kernel function can be applied to the circular convolutional network to synthesize novel views in a variety of occlusion situations. The reconstruction results on extensive datasets validate our model and demonstrate that we can render views for both occluded and nonoccluded scenes.