T-NeRF: Neural Radiance Fields for Traffic Scene

In recent years, Neural Radiance Fields (NeRF) has made significant advancements in novel view synthesis, especially in generating high-quality 3D rendered images. Despite NeRF's unprecedented visual effects, its performance significantly deteriorates when applied to unbounded outdoor traffic scenes, where artifacts often appear during rendering. Additionally, the training time for NeRF is relatively long, especially when processing high-resolution images or complex scenes, with the training process potentially taking several hours to several days. To overcome these problems, this paper presents a custom dataset, NeRF _Traffic, designed to supplement and enhance traffic scene data, thereby enabling better evaluation of NeRF's performance in real-world unbounded traffic scenarios. Based on this dataset, we further propose a new neural radiance field architecture, T-N eRF, for 3D reconstruction of real-world traffic scenes. Experimental results demonstrate that T-NeRF significantly improves rendering quality while greatly reducing training time, providing an effective solution for efficient modeling and rendering of traffic scenes.