http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
DaRF: Few-shot Neural Radiance Field 강화 공동 단안 깊이 적응
송지언(Jiuhn Song),박성훈(Seonghoon Park),곽민섭(Minseop Kwak),백종범(Jongbeom Baek),김승룡,박현희(Hyunhee Park),김낙훈(Nakhoon Kim),이아랑(Arang Lee),전학제(Hakjae Joen) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
We propose a framework called DaRF that combines the strengths of Neural Radiance Field (NeRF) and monocular depth estimation (MDE) to achieve robust reconstruction with only a few real-world images. Existing methods using external priors have limited success, but our approach leverages a powerful MDE network pretrained on large-scale RGB-D datasets. We address the challenges of using MDE with NeRF by incorporating online complementary training. Our framework enhances NeRFs robustness and coherence by imposing the MDE networks geometry prior at both seen and unseen viewpoints. We also tackle the ambiguity issues of monocular depths through patch-wis scale-shift fitting and geometry distillation. Experimental results demonstrate that our framework achieves state-of-the-art performance in both indoor and outdoor real-world datasets, showcasing consistent and reliable results both quantitatively and qualitatively.