Scalable Neural Light Field With Layer Add-ons of Multilayer Perceptron

Light field (LF) is one of 3-D image processing techniques that provides a simple way to generate immersive content. Recently, neural LF (NLF), which incorporates the concept of implicit neural representation into LF, has been introduced, and the difficulties in LF reconstruction have been significantly reduced. Nevertheless, NLF still suffers from the limitations of reusability. To address this issue, this article proposes scalable NLF (S-NLF) that reconstructs LFs of various qualities via a single multilayer perceptron (MLP). S-NLF has a scalable structure that supports low-quality reconstruction with a small MLP, and supports high-quality reconstruction by simply adding auxiliary hidden layers. In addition, this article proposes base layer (BL) sharing to further improve the sample-level efficiency. BL sharing literally shares some part of the hidden layers of different S-NLFs. Experimental results show that S-NLF achieves −40.90% Bjøntegaard delta rate improvement, and BL sharing successfully synthesizes wide-space views with minimal MLPs.