Reproduce “Photometric LiDAR and RGB-D Bundle Adjustment” on OpenResearch

About this paper

Luca Di Giammarino, Emanuele Giacomini, Leonardo Brizi, Omar Salem, Giorgio Grisetti

Sapienza University of Rome developed a unified photometric Bundle Adjustment strategy for global trajectory refinement that works across RGB-D cameras and 3D LiDAR sensors by treating high-resolution LiDAR scans as panoramic images. The approach achieved competitive accuracy on RGB-D (0.006m mean ATE RMSE on ETH3D) and superior robustness and accuracy (approximately 60% improvement) for LiDAR compared to BALM2, with multi-sensor fusion showing enhanced performance.

Key insights

• Modern high-resolution 3D LiDAR scans with reflectivity measurements can be effectively treated as "panoramic images" for direct photometric processing, similar to camera images.
• A unified photometric Bundle Adjustment framework can globally refine trajectories for both RGB-D and LiDAR sensors, leveraging the complementary strengths of each modality (LiDAR's large FOV and robustness, RGB-D's high resolution).
• Multi-sensor fusion, particularly a consecutive strategy (LiDAR for initial alignment, RGB-D for refinement), yields superior convergence and accuracy compared to single-sensor approaches.

Summary by alphaXiv.