Abstract
Though neural radiance fields (NeRF) have demonstrated impressive viewsynthesis results on objects and small bounded regions of space, they struggleon "unbounded" scenes, where the camera may point in any direction and contentmay exist at any distance. In this setting, existing NeRF-like models oftenproduce blurry or low-resolution renderings (due to the unbalanced detail andscale of nearby and distant objects), are slow to train, and may exhibitartifacts due to the inherent ambiguity of the task of reconstructing a largescene from a small set of images. We present an extension of mip-NeRF (a NeRFvariant that addresses sampling and aliasing) that uses a non-linear sceneparameterization, online distillation, and a novel distortion-based regularizerto overcome the challenges presented by unbounded scenes. Our model, which wedub "mip-NeRF 360" as we target scenes in which the camera rotates 360 degreesaround a point, reduces mean-squared error by 54% compared to mip-NeRF, and isable to produce realistic synthesized views and detailed depth maps for highlyintricate, unbounded real-world scenes.