Incremental Convergence Synthesis of Light Rays to Improve Regional Rendering Accuracy in NeRF

作者：Haojie Chen, Zhengyao Bai, Muyuan Cheng · 年份：2024 · DOI：10.1109/iaecst64597.2024.11118174 · 研究领域：Surface Roughness and Optical Measurements、Color Science and Applications、Advanced Image Fusion Techniques

We propose a method to enhance regional reconstruction accuracy for NeRF, which reconstructs the scene from a set of known camera poses and images, and synthesizes novel views of the scene. Building upon a grid-enhanced NeRF architecture, the method optimizes the reconstruction of areas with fewer visible cameras and regions with complex object contours. It also retains fast training speeds, achieving convergence in just 16 minutes with a single GPU, and delivers rendering quality comparable to NeRF models that require several hours of training. In this paper, we investigate the relationships between light ray sampling points during NeRF scene reconstruction. Based on this analysis, we propose a method to strengthen the connections between 3D points in the reconstructed scene, called synthesized rays. This method allows us to fill in the gaps between 3D points and restore local contours and shapes by leveraging surrounding 3D points, thus correcting misrendered 3D points within the region. Experiments show that, under the same 3D perspective warping, the method achieves both faster convergence and better rendering quality on the standard Synthetic dataset.