Neural Density-Distance Fields

The success of neural fields for 3D vision tasks is now indisputable. Following this trend, several methods aiming visual localization (e.g., SLAM) have been proposed to estimate distance or density using fields. However, it difficult achieve high performance by only fields-based such as Neural Radiance Field (NeRF) since they do not provide gradient in most empty regions. On the other hand, field-based Implicit Surface (NeuS) limitations objects’ surface shapes. This paper proposes Distance-Density (NeDDF), a novel representation that reciprocally constrains and We extend field formulation shapes with no explicit boundary surface, fur smoke, which enable conversion from field. Consistent realized both robustness initial values high-quality registration. Furthermore, consistency between allows fast convergence sparse point clouds. Experiments show NeDDF can while providing comparable results NeRF on view synthesis. code available at https://github.com/ueda0319/neddf .