LEGS: Visual Localization Enhanced by 3D Gaussian Splatting.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Accurate six-degree-of-freedom (6-DoF) visual localization is crucial for mapping and navigation.
• Current methods using Novel View Synthesis (NVS) for dataset augmentation often generate redundant or uninformative virtual camera views.
• This limits the effectiveness of bridging sparse real-world captures with dense scene geometry.

Purpose of the Study:

• To introduce LEGS (Visual Localization Enhanced by 3D Gaussian Splatting), a novel framework for trajectory-agnostic synthetic-view augmentation.
• To improve the quality and informativeness of synthetic views for visual localization training datasets.
• To enhance the accuracy and robustness of 6-DoF visual localization.

Main Methods:

• LEGS integrates a coarse 3D lattice with Structure-from-Motion (SfM) camera graphs to propose camera poses.
• A visibility-aware, coverage-driven selection strategy chooses the most informative poses.
• 3D Gaussian Splatting (3DGS) is employed for efficient, scene-specific synthetic view generation.

Main Results:

• LEGS consistently improves 6-DoF pose accuracy and robustness across standard benchmarks and an in-house dataset.
• The framework demonstrates effectiveness particularly in challenging scenarios with sparse camera sampling and co-located viewpoints.
• High-throughput, scene-specific synthesis is achieved within practical computational limits.

Conclusions:

• LEGS offers a significant advancement in synthetic-view augmentation for visual localization.
• The proposed method effectively addresses limitations of naive view sampling in NVS-based approaches.
• LEGS provides a practical and computationally efficient solution for enhancing visual localization systems.