Learning Neural Radiance Field from Quasi-Uniformly Sampled Spherical Image for Immersive Virtual Reality

Le Wang, Shigang Li; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 1166-1180

The neural radiance field (NeRF) is a prominent method of novel view synthesis that is widely applied to many tasks. Recently, the NeRF method, which was originally developed for perspective images, has been extended to 360-degree omnidirectional images, which may make it easier to perform immersive virtual reality tasks. However, in existing NeRF methods, omnidirectional images are typically represented as rectangular images. Equirectangular image representation is popular but is also known for distortion, especially as it approaches the poles. In this paper, we propose a method for learning a neural radiance field from geodesic dome division-based discrete spherical images. First, an input equirectangular image is mapped to a unit sphere. Then, the mapped spherical image is sampled on a unit sphere on the basis of geodesic dome division. Next, the obtained discrete spherical image is used to train an NeRF network. Finally, comparative experiments are conducted for the equirectangular image representation and geodesic dome division discrete spherical image representation. The experimental results demonstrate that the proposed method outperforms existing equirectangular image representation methods.