ReefMapGS: Enabling Large-Scale Underwater Reconstruction by Closing the Loop Between Multimodal SLAM and Gaussian Splatting
New AI framework creates high-fidelity 3D maps of coral reefs without needing COLMAP, using multimodal sensor fusion.
A team of researchers from MIT and the Woods Hole Oceanographic Institution (WHOI) has unveiled ReefMapGS, a groundbreaking AI framework that enables large-scale, high-quality 3D reconstruction of underwater environments. The system solves a critical bottleneck in field robotics by eliminating the need for COLMAP, a computationally intensive structure-from-motion tool traditionally required for accurate camera pose estimation. Instead, ReefMapGS directly leverages multimodal sensor data—acoustic, inertial, pressure, and visual—from autonomous underwater vehicles (AUVs) to feed a pose-graph optimization-based SLAM system. This provides the precise camera poses needed for 3D Gaussian Splatting, a powerful neural rendering technique, while also quantifying uncertainty.
ReefMapGS operates through an incremental reconstruction process. It starts by building an initial 3D Gaussian Splatting model from a region of high certainty and then progressively expands to incorporate the entire scene. The framework interleaves local tracking of new image observations with optimization of the underlying 3D scene representation. Crucially, the refined camera poses from the 3DGS optimization are fed back into the SLAM pose-graph, creating a closed loop that globally optimizes the vehicle's entire trajectory. This bidirectional feedback results in both a more accurate map and more reliable navigation data.
The team validated ReefMapGS on two complex underwater reef sites, demonstrating its ability to perform detailed 3D reconstruction without COLMAP. The system also achieved more accurate global pose estimation for AUVs over long survey trajectories spanning up to 700 meters. This represents a significant advancement for marine science, underwater inspection, and archaeology, where creating precise, large-scale 3D models in challenging, feature-poor environments has been a major technical hurdle.
- Eliminates dependency on COLMAP for camera pose estimation, using real-time multimodal SLAM instead.
- Creates a closed-loop system where 3DGS refines poses fed back to SLAM, improving both map and trajectory accuracy.
- Successfully reconstructed complex reef geometry and tracked AUVs over 700m trajectories in field tests.
Why It Matters
Enables autonomous underwater robots to create accurate, large-scale 3D maps for marine conservation and infrastructure inspection in real-time.