EditableNeRF: Editing Topologically Varying Neural Radiance Fields by Key Points.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Neural radiance fields (NeRF) excel at photo-realistic novel-view synthesis.
• Editing dynamic scenes modeled by NeRF remains a significant challenge, particularly concerning topological variations.

Purpose of the Study:

• To develop an editable neural radiance field system for intuitive scene editing.
• To enable end-users to modify dynamic scenes, including topological changes, using simple interactions.

Main Methods:

• A novel scene analysis method detects and initializes surface key points considering scene dynamics.
• A weighted key points strategy optimizes joint key points and weights for modeling dynamic topology.
• The network is trained automatically from single-camera image sequences.

Main Results:

• The proposed method allows intuitive multi-dimensional (up to 3D) editing of dynamic scenes.
• High-quality scene editing results were achieved on various dynamic scenes.
• The system successfully generates novel scene views not present in the input data.

Conclusions:

• Editable neural radiance fields offer a powerful tool for user-driven dynamic scene manipulation.
• The approach effectively handles topological changes in dynamic scenes.
• This method outperforms existing state-of-the-art techniques in dynamic scene editing.