Selective sampling-based radiance field learning for wide-angle view synthesis
Novel view synthesis aims to reconstruct a scene and generate previously unseen viewpoints from given images. While Neural Radiance Fields have achieved strong results, prior methods typically rely on diverse multi-view inputs and accurate poses, which is difficult to satisfy in large-scale scenes captured along long-term trajectories from frontal views. In this paper, we propose a framework that enables wide-angle view synthesis using only frontal-view images acquired along long-term trajectories. Our key insight is that the spatial continuity inherent in long-term trajectories can be exploited to progressively accumulate scene coverage, allowing radiance fields to represent wide viewing angles without omnidirectional inputs. In particular, we propose a selective sampling-based spatial contraction strategy that allows each radiance field to learn from extended trajectories while prioritizing recently observed frames, enhancing spatial coverage and reconstruction stability in unbounded scenes. Moreover, we introduce a proximity-aware blending mechanism that adaptively assigns rendering weights according to spatial distances between camera rays and radiance-field centers, yielding seamless transitions and consistent appearance across multiple fields. Extensive experiments on two public datasets demonstrate the efficiency and effectiveness of our approach.
Authors
- Katsuya Hotta (ORCID: https://orcid.org/0000-0002-8148-1445)
- Yoshihiro Hagihara
- Kansei Ogura
Institutions
- Iwate University (JP)
Publication Details
- Journal
- Signal Image and Video Processing
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1007/s11760-026-05855-9
- Primary Topic
- 3D Shape Modeling and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00