D-NeRF turns selfie videos from your phone into free-viewpoint portraits which we call nerfies.
Our approach augments Neural Radiance Fields by optimizing an additional continuous volumetric deformation field that warps each observed point into a canonical 5D NeRF. These deformation fields are prone to local minima so the project also proposes additional constraints for more robust optimization.
The project can create fun visual effects like the dolly zoom, solve for matting, interpolate between observations and re-render video from novel viewpoints.
More details and figures can be found on the project page.