LEGS: Visual Localization Enhanced by 3D Gaussian Splatting

TL;DR

LEGS improves visual localization by using 3D Gaussian Splatting to generate better synthetic camera views for training.

Contribution

LEGS introduces a trajectory-agnostic framework using 3D Gaussian Splatting for efficient and effective synthetic-view augmentation.

Findings

LEGS improves pose accuracy and robustness on standard benchmarks.

The method performs well in scenarios with sparse sampling and co-located viewpoints.

3D Gaussian Splatting enables high-throughput scene-specific synthesis.

Abstract

Accurate six-degree-of-freedom (6-DoF) visual localization is a fundamental component for modern mapping and navigation. While recent data-centric approaches have leveraged Novel View Synthesis (NVS) to augment training datasets, these methods typically rely on uniform grid-based sampling of virtual cameras. Such naive placement often yields redundant or weakly informative views, failing to effectively bridge the gap between sparse, unordered captures and dense scene geometry. To address these challenges, we present LEGS (Visual Localization Enhanced by 3D Gaussian Splatting), a trajectory-agnostic synthetic-view augmentation framework. LEGS constructs a joint set of 6-DoF camera pose proposals by integrating a coarse 3D lattice with the Structure-from-Motion (SfM) camera graph, followed by a visibility-aware, coverage-driven selection strategy. By utilizing 3D Gaussian Splatting (3DGS), our framework enables high-throughput, scene-specific synthesis within practical computational budgets. Experiments on standard benchmarks and an in-house dataset demonstrate that LEGS consistently improves pose accuracy and robustness, particularly in scenarios characterized by sparse sampling and co-located viewpoints.

Figures