EditableNeRF: Editing Topologically Varying Neural Radiance Fields by Key Points

Knocking on each piano key (a) Input Sequence EditableNeRF Training Editing by Key Points (b) Reconstruction (c) Editing Results Novel view (key point) Playing a piece of music Sliding on the piano keys EditableNeRF Training Editing by Key Points Shaking and lifting either cup (one moves, the other stands) Novel view (2 key points) Freely moving for two dice cups Figure 1. Taking an image sequence (a) as input, EditableNeRF is trained fully automatically to reconstruct the captured scene (b) and can handle topological changes. After training, end-users are able to edit the scene (c) by controlling the automatically picked-out key points (circled in green in (b)). Our method enables multi-dimensional editing and can generate novel scenes that are unseen during training.