Geometric Neural Process Fields
Keywords: Implicit Neural Fields, Neural Processes, Generalization
Abstract: This paper focuses on Implicit Neural Representation (INR) generalization, where models need to efficiently adapt to new signals with few observations. Specifically, for radiance field generalization, we propose Geometric Neural Processes (GeomNP) for probabilistic neural radiance field to explicitly capture uncertainty. We formulate INR generalization in a probabilistic manner, which incorporates uncertainty and directly infers the INR function distributions on limited context observations. To alleviate the information misalignment between the 2D context image and 3D discrete points in INR generalization, we introduce a set of geometric bases. The geometric bases learn to provide 3D structure information for inferring the INR function distributions. Based on the geometric bases, we model GeomNP with hierarchical latent variables. The latent variables integrate 3D information and modulate INR functions in different spatial levels, leading to better generalization of new scenes. Despite being designed for 3D tasks, the proposed method can seamlessly apply to 2D INR generalization problems. Experiments on novel view synthesis of 3D ShapeNet and DTU scenes, as well as 2D image regression, demonstrate the effectiveness of our method.
Supplementary Material: zip
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: 7536
Loading