Geometry and Uncertainty Guided Active View Selection and Synthetic View Augmentation for 3D Gaussian Splatting

Abstract

High-quality 3D scene reconstruction depends strongly on the number and spatial distribution of input views. When available images are sparse or unevenly distributed, re-construction quality and generalization to unseen viewpoints can degrade substantially. To address this, this paper presents an active view selection framework for data-efficient 3D reconstruction based on 3D Gaussian Splatting. The proposed method combines pose-aware data partitioning with a lightweight acquisition strategy that integrates geometric coverage and model uncertainty. Geometric coverage is used to favor spatially informative view-points, while uncertainty is estimated through pose jittering and render variance. The framework is evaluated in two settings: real-view expansion from a candidate pool and NeRF-guided synthetic augmentation from sampled novel poses. Experimental results on held-out test views show that the proposed strategy consistently improves reconstruction quality as informative views are added. In particular, PSNR and SSIM generally increase while decreases LPIPS across both branches. The results further indicate that real-view active selection provides the strongest and most reliable gains, while actively selected synthetic views can also be effective as a complementary augmentation strategy. Overall, the proposed framework offers a practical and scientifically grounded approach for efficient view acquisition in neural scene reconstruction.