Active Reconstruction of Radiance Fields with Expanded Silhouette Refinement.

Abstract

Radiance field representations have achieved notable success in novel view synthesis and 3D geometry reconstruction. Nevertheless, the reconstruction quality is highly sensitive to both the number of input images and the extent of scene coverage, and it often requires manual effort to capture well-conditioned image collections. To reduce acquisition cost while preserving reconstruction fidelity, active view selection methods aim to determine the next-best-view (NBV) that maximizes information gain at each iteration. However, existing approaches typically restrict the camera pose search space during NBV optimization, which can lead to sub-optimal view selection and limited reconstruction quality. In this paper, we expand the NBV candidates sampling space to identify informative image collections that jointly balance local surface detail and global geometric completeness. We propose a two-stage Silhouette Refinement Candidate Generation (SRCG) framework that leverages prior knowledge of well-composed viewpoints to generate candidate views with improved object coverage. We optimize the coarse camera candidates in a differentiable manner through volume rendering, guided by a proposed Expanded Silhouette Loss that encourages effective silhouette capture. Extensive evaluations on synthetic and real-world benchmarks using Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) backbones demonstrate that the proposed approach produces candidate views with superior surface coverage and pixel density and consistently improves novel view synthesis across diverse object categories.