Overview
HyperNeRF addresses the challenge of modeling topological changes in dynamic scenes using Neural Radiance Fields (NeRF). It lifts NeRFs into a higher-dimensional space, representing the 5D radiance field for each input image as a slice through this hyper-space. Inspired by level set methods, it models changes in scene topology by providing a NeRF with a higher-dimensional input. The architecture extends Nerfies by conditioning the template NeRF on additional higher-dimensional coordinates, effectively creating an 'ambient slicing surface'. This enables the interpolation and novel-view synthesis of scenes with topological variations, outperforming existing methods. It improves average error rates by 4.1% for interpolation and 8.6% for novel-view synthesis, as measured by LPIPS, compared to Nerfies.