Multiple Level of Details Neural Implicit Surface Representation for Unconstrained Viewpoint Rendering

Zhihua Chen, Yuhang Li, Lei Dai, Ping Li, Lei Zhu, Bin Sheng

Abstract: To mitigate artifacts and performance degradation in neural implicit representation visualization for unconstrained viewpoint rendering, we propose a feature voxel grid-based neural representation architecture. This approach flexibly encodes the implicit surface with multiple levels of detail, facilitating high-quality rendering with dynamic switching between detail levels. Additionally, we implement a range-limited strategy to concentrate on sampling valid areas while excluding undefined areas. Our results in unconstrained viewpoint scenarios demonstrate the effectiveness of our method. This work extends the capabilities of neural implicit representations, broadening their potential applications beyond previously defined limitations.