near all volumetric representations such as Nerf enable photorealistic rendering of complex large-scale scenes but are to compute intensive for real-time rendering existing approaches for Accelerated neural volumetric rendering rapidly consume memory as seen resolution increases we present a memory efficient Radiance field representation which we call Murph that enables real-time high resolution rendering of large-scale scenes when rendered on a more powerful GPU our method can run at frame rates as high as 60 FPS for a given memory budget Murph can represent significantly more fine details than state-of-the-art real-time Nerf rendering techniques such as snerg and Murph occupies
considerably less video memory when representing scenes at the same resolution as snerg while incurring only a minimal loss in quality Murph represents a scene as a 3D volumetric field characterized by a sparse low resolution feature grid and three high resolution feature planes we query the representation at any 3D Point by gathering features from the grid and the projection of the point onto the three planes these features are summed together to get the volume density and diffuse color for that point along with a feature Vector that encodes view dependent appearance both the high resolution low rank
and low resolution high rank components of our model are crucial for representing detailed scenes for example if we remove the low resolution feature grid and only represent the scene with feature planes we can observe significant rendering artifacts by baking a volumetric representation of the scene we preserve higher rendering quality than techniques such as mobile Nerf which bakes a Nerf into a polygon mesh representation the Simplicity of our Murph representation leads to lower memory bandwidth requirements resulting in much faster rendering than techniques such as instant NGP Murph is effective for real-time volumetric rendering across a wide
variety of large-scale scenes