Abstract

Neural Radiance Fields (NeRF) have become more and more popular in the field of 3D vision. Although they perform very well, they require dense input views as supervision. Given sparse input images, the quality of NeRF's reconstruction drops significantly due to the radiance-ambiguity problem. We address this problem by leveraging priors from already available multi-view stereo (MVS) models. We propose to regularize NeRF's optimization with the MVS probability volumes. In addition, NeRF can guide the depth hypothesis sampling to further boost the performance of MVS methods. Given only sparse input views, experiments show that our method not only outperforms generic neural rendering models by a large margin, but also significantly increases the reconstruction quality of MVS models.