T(G)V-NeRF: A Strong Baseline in Regularized Neural Radiance Fields with Few Training Views

Implicit representations such as Neural Radiance Fields (NeRFs) have become a de facto standard in the field of novel view synthesis for 3D scenes. However, their stunning results typically imply the use of dozens of training images, with their corresponding cameras well localized in the scene. This paper studies new, total variation-based regularization approaches to train NeRFs in the context of very few (less than 10) training images. It leverages the NeRF back-propagation algorithm to evaluate first-order and second-order derivatives terms on the inferred depth map to enforce smoothness on the scene underlying surfaces. Through state-of-the-art performance on standard real-images benchmarks, we show that the proposed methods, coined as TV-NeRF and TGV-NeRF, make strong baselines in novel view synthesis with few training views.