BavsNeRF: Batch Active View Selection for Neural Radiance Field Using Scene Uncertainty

You Li, Rui Li, Shengjun Tang, Yanjie Gu

Keywords: NeRF, Active View Selection, Uncertainty, 3D reconstruction

TL;DR: Utilizing neural radiance fields (NeRF) to compute scene uncertainty, our framework guides batch active view selection tasks.

Abstract: Active view selection is crucial for Neural Radiance Fields (NeRF) modeling in scenarios with limited number of posed images. Existing methods to select the views are either heuristic or computationally demanding. To address this, we propose a novel framework, BavsNeRF, to guide our view selection for NeRF modeling using scene uncertainty. We first establish an uncertainty estimation model of the entire scene based on an initial NeRF model. With this, we guide new perceptions by incorporating an batch active view selection policy, enabling the entire view selection procedure within a single iteration. In this way, the quality of novel view synthesis can be enhanced by incorporating images from selected viewpoints containing informative data. Experiments on both synthetic and real-world datasets demonstrate that the proposed method can identify informative new viewpoints, leading to more accurate scene reconstruction compared to baseline and state-of-the-art methods.

Submission Number: 7