Reproduce “Uncertainty-Aware Visual-Inertial SLAM with Volumetric Occupancy Mapping” on OpenResearch

About this paper

Jaehyung Jung, Simon Boche, Sebastián Barbas Laina, Stefan Leutenegger

Researchers from the Smart Robotics Lab at TUM developed a visual-inertial SLAM system that tightly integrates learned, pixel-wise depth uncertainties from deep neural networks into both volumetric occupancy mapping and factor graph optimization. This approach achieved state-of-the-art localization accuracy on EuRoC (0.041m ATE RMSE causal) and the Hilti-Oxford dataset, while producing more accurate and complete dense maps than prior methods.

Key insights

• Leveraging pixel-wise, learned depth uncertainty from deep networks significantly enhances the statistical consistency of both volumetric maps and the accuracy of state estimation, improving robustness in challenging environments.
• Optimally fusing complementary depth information from static stereo and motion stereo (MVS) networks, guided by their respective learned uncertainties, yields superior depth quality and completeness across various ranges.
• A tightly coupled framework where dense map information, weighted by its uncertainty, directly influences the core SLAM factor graph optimization leads to globally consistent and highly accurate localization and mapping.

Summary by alphaXiv.