Segmentation-Guided Neural Radiance Fields for Foreground Object 3D Reconstruction

Abstract

In recent years, Neural Radiance Fields (NeRF) is attracting attention for its excellent performance in reconstructing 3D scenes from 2D images by capturing volumetric scene representations and radiance properties. However, NeRF is challenged with the concurrent rendering of both foreground and background components, which raises computational complexity issues for scene reconstruction. In response to these challenges, our approach combines image segmentation to delineate objects with precision. By focusing NeRF exclusively on foreground objects during training, we optimize its rendering capacity. This integration achieves contextually accurate 3D reconstructions, demonstrated with high-quality results. Through qualitative evaluation of various data, we show that removing the background improves 3D reconstruction precision as well as computation speed.