GRF LEARNING A GENERAL RADIANCE FIELD FOR 3D SCENE REPRESENTATION AND RENDERING
Published:
NeRF: shape and appearance only optimized by RGB values, lack grneral geometric patterns
- extracting 2D features for every light from input 2D images
- repeoject to 3D point
- selecting and aggregating features, implicitly considering visual occlusions
- rendering
- well, that is not a fair comparation: CNN improves performance by a large margin
- more params!
- but, adapt to novel scene fast!