Well-NeRF: Ensuring Well-Posed Neural Radiance Fields via View Frustum and Shadow Zone Based Regularization

Geunho Kim, Jinwook Paeng, Jin Hong, Yoojin Han, Junseok Kwon

Keywords: Few-shot NeRF, Ill-posed problem, Artifacts removal, View frustum, Inside opaque, Boundary condition, Near-far threshold, Integrated model

TL;DR: We demonstrate that applying boundary conditions on the learning space of the NeRF model, based on an analysis of position-inferable regions, is highly effective in mitigating the sparse input problem.

Abstract: Neural Radiation Field (NeRF) often produces many artifacts with sparse inputs. These artifacts are primarily caused by learning in regions where position inference is not feasible. We assume that the main cause of this problem is the incorrect setting of boundary conditions in the learning space. To address this issue, we propose a new regularization method based on two key assumptions: (1) the position of density and color cannot be inferred in regions where the view frustum does not intersect, and (2) information inside opaque surfaces cannot be observed and inferred, and thus cannot contribute to the rendering of the image. Our method aims to transform the NeRF model into a well-posed problem by regularizing learning in regions where position inference is not possible, allowing the network to converge meaningfully. Our approach does not require scene-specific optimization and focuses on regions where position inference is not possible, thereby avoiding degradation of model performance in main regions. Experimental results demonstrate the effectiveness of our method in addressing the sparse input problem, showing outstanding performance on the Blender synthetic datasets. Our method is designed to integrate seamlessly with existing techniques, providing an effective solution for sparse input scenarios and offering a foundational approach that serves as the first clue in addressing sparse input problems.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9236