A NeRF-based technique combined depth-guided filtering and view enhanced module for large-scale scene reconstruction

Abstract

Efficient and high-quality rendering of large-scale geographic scenes is crucial in the field of virtual geographic environment. Neural Radiance Field (NeRF), as a novel and cutting-edge view synthesis technique, has the great simplicity of process and the higher fidelity of rendering effect. However, when applied to large scenes, constrained by the network performance and the lack of 3D geometric model, NeRF still needs to be improved in terms of the visual sharpness of modeling results and the recognition ability of geometric features. In this paper, a depth-guided filtering method is designed for punishing the noise and visual artifacts in radiance field. In addition, a view enhanced module is proposed, which fuses adjacent high-quality reference views to greatly improve the clarity of rendered images. Moreover, on two public large-scale geographic datasets and our constructed campus dataset, extensive experiments have shown that our method not only achieves better high-quality reconstruction results than traditional explicit modeling methods, but also exceeds the common implicit modeling methods 6.91 % at most in reconstruction accuracy.