Enhancing NetVLAD based localization with Gaussian Splat

We propose enhancing NetVLAD-based localization by generating additional virtual images using 3D Gaussian splatting. Gaussian splatting models a scene with 3D Gaussian kernels, allowing novel views to be rendered from arbitrary positions. By integrating these virtual images into the reference database, we increase the density of visual information, improving localization accuracy. Using the 7Scenes dataset, we demonstrate that augmenting the dataset with virtual images significantly enhances localization performance, especially in environments with sparse data. This approach offers a robust solution for visual localization in applications with limited real-world data collection. © 2025 IEEE.