GSDB: A Lightweight Database for Gaussian Splatting Map-based Visual Localization Leveraging Edge-aware and Quality-guided View Filtering

This paper presents the Gaussian splatting database (GSDB), a database construction pipeline for Gaussian splatting map-based visual localization. GSDB consists of three stages: (1) View filtering using pixel-wise gradient variance and frustum constraints to remove visually uninformative views; (2) Covisibility graph construction by analyzing spatial overlaps between frustums of among viewpoints; and (3) Viewpoint ranking using perceptual quality (BRISQUE) and structural reliability (Edge-Aware Uncertainty Concentration, EAUC) to select representative views from each covisibility group. This process ensures that only the most informative and stable views are retained in the final database. The proposed GSDB framework is compatible with Gaussian Splatting-based maps and does not require additional geometry or reference supervision. It enables significant database compression while maintaining localization accuracy. Experimental results on the Replica dataset show that GSDB effectively reduces database redundancy, improves retrieval efficiency, and results in up to about 10% faster localization process than conventional baselines, while compressing the database size by up to 95%.

Our GSDB pipeline converts a dense Gaussian Splatting map into a compact, retrieval-friendly database in three main stages. First, we perform view filtering: all training viewpoints are rendered, and we discard views that are visually uninformative using pixel-wise gradient variance, then further reject views whose frustums observe only a small number of high-opacity Gaussians, ensuring that each retained image sees meaningful 3D structure. Second, from the remaining views we build a covisibility graph by connecting view pairs that share a sufficiently large set of visible Gaussians; each connected component of this graph defines a covisibility group that observes a similar region of the scene. Third, within each group we apply viewpoint ranking using two complementary scores: a no-reference perceptual quality metric (BRISQUE) to penalize blurred or artifact-ridden renderings, and our Edge-Aware Uncertainty Concentration (EUAC), which measures how uncertainty derived from Fisher Information is distributed along edges detected by DiffusionEdge. By selecting the top-ranked representative view from every covisibility group, GSDB preserves the most informative and structurally reliable viewpoints, achieving over 90% database compression while maintaining localization accuracy and improving overall retrieval and pose-estimation latency.