Abstract

We introduce a novel versatile approach to enhance the quality of grid-based neural scene representations. Grid-based scene representations model a scene by storing density and color features at discrete 3D points, which offers faster training and rendering than purely implicit methods such as NeRF. However, they require high-resolution grids to achieve high-quality outputs, leading to a significant increase in memory usage. Common standard grids with uniform voxel sizes do not account for the varying complexity of different regions within a scene. This is particularly evident when a highly detailed region or object is present, while the rest of the scene is less significant or simply empty. To address this we introduce a novel approach based on frequency domain transformations for finding the key regions in the scene and then utilize a 2-level hierarchy of grids to allocate more resources to more detailed regions. We also created a more efficient version of this concept, that adapts to a compact grid representation, namely TensoRF, which also works for very few training samples.