Clean-NeRF: Defogging using Ray Statistics Prior in Natural NeRFs
Keywords: NeRF, novel view synthesis
Abstract: State-of-the-art Neural Radiance Fields (NeRFs) still struggle in novel view synthesis for complex scenes, producing inconsistent geometry among multi-view observations, which is manifested into foggy ``floaters'' typically found hovering within the volumetric representation. This paper introduces Clean-NeRF to improve NeRF reconstruction quality by directly addressing the geometry inconsistency problem. Analogous to natural image statistics, we first perform empirical studies on NeRF ray profiles to derive the {\em natural ray statistics prior}, which is employed in our novel ray rectification transformer capable of limiting the density only to have positive values in applicable regions, typically around the first intersection between the ray and object surface. Moreover, Clean-NeRF automatically detects and models view-dependent appearances to prevent them from interfering with density estimation. Codes will be released.
Primary Area: representation learning for 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/2024/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: 482
Loading