The research addresses the issue of active view selection and uncertainty quantification in Radiance Fields. The researchers propose a method that utilizes Fisher Information to efficiently quantify observed information within Radiance Fields without the need for ground truth data. This is used for selecting the next best view and quantifying pixel-wise uncertainty. The method overcomes existing limitations on model architecture and effectiveness, achieving state-of-the-art results in both view selection and uncertainty quantification. The method, when used with the 3D Gaussian Splatting backend, could perform view selections at 70 fps.
Publication date: 29 Nov 2023
Project Page: https://jiangwenpl.github.io/FisherRF/
Paper: https://arxiv.org/pdf/2311.17874
