SID-NERF: Few-Shot Nerf Based on Scene Information Distribution

Abstract

The novel view synthesis from a limited set of images is a significant research focus. Traditional NeRF methods, relying mainly on color supervision, struggle with accurate scene geometry reconstruction when faced with sparse input images, leading to suboptimal rendering. We propose a Few-shot NeRF Based on Scene Information Distribution(Sid-NeRF) to address this by integrating geometric and color supervision, enhancing the model’s understanding of scene geometry. We also implement a data selector during training to identify and utilize the most accurate geometric data, thus improving training efficiency. Additionally, a residual module is introduced to counteract any optimization biases from the selector. Our method was tested on three datasets and showed excellent performance in various environments with limited images. Notably, compared to other novel view synthesis methods based on fewer views, our method does not require any prior knowledge and thus does not incur additional computational and storage costs.