Urban-level three-dimensional reconstruction for modern applications demands high rendering fidelity while minimizing computational costs. The advent of Neural Radiance Fields (NeRF) has enhanced 3D reconstruction, yet it exhibits artifacts under multiple viewpoints. In this paper, we propose a new NeRF framework method to address these issues. Our method uses image content and pose data to iteratively plan the next best view. A crucial aspect of this method involves uncertainty estimation, guiding the selection of views with maximum information gain from a candidate set. This iterative process enhances rendering quality over time. Simultaneously, we introduce the Vonoroi diagram and threshold sampling together with flight classifier to boost the efficiency, while keep the original NeRF network intact. It can serve as a plug-in tool to assist in better rendering, outperforming baselines and similar prior works.

翻译：面向现代应用的城市场景三维重建需要在保证高渲染保真度的同时最小化计算成本。神经辐射场（NeRF）的出现提升了三维重建的质量，但在多视角下仍存在伪影。本文提出一种新的NeRF框架方法以解决这些问题。我们的方法利用图像内容与位姿数据，通过迭代式规划选择下一个最优视角。该方法的一个关键环节在于不确定性估计，其通过从候选视角集合中选择信息增益最大的视角来指导视角选取。这一迭代过程可随时间逐步提升渲染质量。同时，我们引入Voronoi图与阈值采样方法，并结合飞行分类器以提升效率，同时保持原始NeRF网络结构不变。该方法可作为即插即用工具辅助提升渲染效果，其性能优于基线方法及同类已有工作。