| Title |
Boosting Few-shot Radiance Fields with Monocular Depth Estimation |
| Authors |
박성훈(Seonghoon Park) ; 조은지(Eunji Cho) ; 정호민(Homin Jung) ; 안종식(Jongsik Ahn) ; 김승룡(Seungryong Kim) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.99 |
| Keywords |
Novel view synthesis; NeRF; Monocular depth estimation |
| Abstract |
Neural radiance field (NeRF) excels in novel view synthesis and 3D reconstruction but struggles when known viewpoints are few. Existing solutions using external priors are limited to specific scenes or datasets. Using monocular depth estimation (MDE) networks pretrained on large-scale RGB-D datasets can help, but MDE introduces ambiguity issues. We propose a novel framework, DaRF, which combines NeRF and MDE through online complementary training, enabling robust NeRF reconstruction from a few real-world images. Our framework applies MDE's strong geometry prior to NeRF at both seen and unseen viewpoints, enhancing robustness and coherence. We address MDE’s ambiguity through patch-wise scale-shift fitting and geometry distillation, aligning MDE depths with NeRF geometry. Experiments show our framework achieves state-of-the-art results, performing consistently well on indoor and outdoor real-world datasets. |