Warp Consistent Neural Radiance Fields for Sparse Novel View Synthesis

Yexing Xu, Minglin Chen, Longguang Wang, Ye Zhang, Yulan Guo

Abstract: Neural Radiance Field (NeRF) has achieved photorealistic rendering in novel view synthesis, but its performance significantly degrades when input images are sparse. Few images impose limited constraints for the reasoning of accurate geometry and appearance, leading to 3D inconsistency in novel view synthesis. In this work, we present a novel regularization approach (namely, Warp-NeRF) to enhance NeRF with sparse input images via warp consistent constraints. Specifically, we employ image warping to establish mapping relations and regularize the consistency in latent space to improve the quality of rendered images. In addition, we propose several filtering operations to exclude the effect of outliers and make our regularization more robust. Comprehensive experiments show that the proposed method can render 3D consistent novel views with only 3 input images and achieves state-of-the-art performance on both the LLFF and Shiny datasets.