Square-LFRF: high-speed light field 3D source generation using anti-aliasing virtual views synthesis based on neural radiance fields

Light field 3D display based on integral imaging allows for glasses-free 3D perceptions by providing parallaxes in two dimensions. The light field 3D sources serve as the data source, which are merged from several parallax images. However, with the sharp increase in the number of viewpoints, both the speed and quality of generating light field 3D sources are limited. To address these issues, we propose a Square-LFRF method that enables generation of light field 3D sources with high speed using high-fidelity and anti-aliasing virtual views synthesis based on neural radiance field (NeRF). The arbitrary sparse viewpoints are used as the input to reconstruct the light field of 3D scenes. We employ a square pyramid frustum casting and recalibrate the sampling integrated position encoding to enhance the quality of virtual views. Compared to the conventional Mip-NeRF, Square-LFRF reduces average relative error rates by 8% on the Blender dataset. The network size and training time are reduced to 72.3 MB and 10 minutes through cubic projection, respectively. To further speed up the generation of light field 3D sources, we propose a pixel culling method to eliminate the rendering of redundant rays. Experimental results show that Square-LFRF can generate light field 3D sources 70% faster than the conventional NeRF-based 3D source generation method of cutoff-NeRF.