MrNeRF (@janusch_patas)
2025-01-15 | โค๏ธ 166 | ๐ 28
VINGS-Mono: Visual-Inertial Gaussian Splatting Monocular SLAM in Large Scenes
Note: Mesh is created from Kitty dataset.
Contributions: โข We are the first monocular (inertial) GS-based SLAM system capable of operating outdoors and supporting kilometer-scale urban scenes.
โข We propose a 2D Gaussian Map module, including a sample rasterizer, score manager, and single-to-multi pose refinement, ensuring accurate localization and high-quality Gaussian maps in real time.
โข We introduce a GS-based loop detection method, along with an efficient approach that corrects tens of millions of Gaussian attributes in a single operation upon loop detection, effectively eliminating accumulated errors and ensuring global map consistency.
โข Comprehensive experiments on different scenes (indoor environments, aerial drone view, and driving scenes) demonstrate that VINGS-Mono outperforms existing approaches in rendering and localization performance. We also developed a mobile app and conducted real-world experiments to demonstrate the practical reliability of our method.
๋ฏธ๋์ด
๐ Related
Auto-generated - needs manual review