<scp>LCA</scp> ‐ <scp>NeRF</scp> : Large‐Scale Scene Neural Rendering With Clustered Appearance Embedding

ABSTRACT Recently, Neural Radiance Fields (NeRF) have been used for urban‐scale scenes with potentially infinite scales. Prior work often necessitates training times of several tens of hours in large‐scale scenes. Although fast‐converging NeRF variants have been applied to urban‐scale scenes, they struggle to capture sufficient details and prove challenging in handling dramatic lighting variations in practical usage. It could be argued that these limitations arise from two key factors: limited model capacity and insufficient attention to the geometric distribution of nonlinear light sources within the scene. To address these challenges, we propose a novel approach for modeling large‐scale scenes. Building upon the foundation of fast‐converging NeRF variants, we incorporate a block‐based strategy to reduce training costs and capture additional scene details. Furthermore, we introduce clustered appearance embeddings to model the nonlinear lighting present in space, enabling smoother transitions in radiance during practical use. We evaluated our approach against other methods for large scenes on the Mill19 and Urbanscene3D datasets, surpassing the state‐of‐the‐art methods and converging within a few hours.