RGU-AVS: Residual-Driven Gaussian Uncertainty for Pool-Based Active View Selection in 3D Gaussian Splatting

Abstract

Pool-based active view selection seeks to identify an informative subset of training images from a pre-acquired image pool, thereby reducing redundant observations of 3D Gaussian Splatting (3DGS). Existing methods mainly estimate candidate-view utility from predictive uncertainty, parameter sensitivity, or geometric coverage, while overlooking the photometric reconstruction errors already available in the candidate image pool. Therefore, we propose a residual-driven Gaussian uncertainty framework (RGU-AVS), which attributes photometric residuals directly to contributing Gaussian primitives based on their alpha-blending contributions and propagates the resulting Gaussian-level uncertainty back to candidate viewpoints for view selection. Experiments on four public datasets demonstrate that RGU-AVS achieves competitive overall rendering quality under identical training-view budgets. These results indicate that residual-to-Gaussian attribution provides a highly effective criterion for identifying informative training views in pool-based 3DGS reconstruction.