Effective Prior Regularized Sparse Learning

Abstract

Neural radiance fields (NeRF) have the ability of synthesizing novel views from sets of input images, which has attracted a great deal of interest in recent years. Typical methods require tens of images for view synthesis, which limits the potential applications of NeRF. In this paper, a novel framework is proposed for view synthesis in a sparse setting by tactically imposing a regularization using prior information extracted from a pretrained network. We design a network model that trains a prior field as well as a color field simultaneously, and the network integrates such prior knowledge for better novel view synthesis. Experiments on two benchmark datasets have demonstrated the effectiveness and robustness of our method and that our framework is adaptable to other existing methods for synthesizing better quality outputs in a sparse setting.