Abstract
This paper proposes a photorealistic real-time dense 3D mapping system that utilizes a learning-based image enhancement method and mesh-based map representation. Due to the characteristics of the underwater environment, where problems such as hazing and low contrast occur, it is hard to apply conventional simultaneous localization and mapping (SLAM) methods. Furthermore, for sensitive tasks like inspecting cracks, photorealistic mapping is very important. However, the behavior of Autonomous Underwater Vehicle (AUV) is computationally constrained. In this paper, we utilize a neural network-based image enhancement method to improve pose estimation and mapping quality and apply a sliding window-based mesh expansion method to enable lightweight, fast, and photorealistic mapping. To validate our results, we utilize real-world and indoor synthetic datasets. We performed qualitative validation with the real-world dataset and quantitative validation by modeling images from the indoor synthetic dataset as underwater scenes.
Introduction
Underwater environments present a "triple threat" to autonomous navigation and inspection. Light scattering and absorption cause heavy hazing, color distortion, and low contrast, which often lead to the failure of standard Simultaneous Localization and Mapping (SLAM) algorithms. While Autonomous Underwater Vehicles (AUVs) are critical for inspecting submerged infrastructure, they are hampered by visual degradation, limited onboard processing power, and the lack of interpretability in standard map representations like point clouds or voxels.
Proposed Methodology
The researchers developed a three-stage pipeline designed to overcome these environmental and computational hurdles.
Neural Image Enhancement
The system integrates Joint-ID, a transformer-based neural network that performs image enhancement and depth estimation simultaneously. By restoring color and contrast to raw, hazy footage, the system provides a cleaner signal for downstream tasks.
Robust Localization and Mesh Mapping
The enhanced images are processed using ORB-SLAM2. The improved visual clarity allows the algorithm to detect and track 16.1% more feature points, ensuring stable pose estimation even during high-speed rotations.
For mapping, the system generates a 2D mesh using Delaunay triangulation on feature points, which are then projected into 3D using the estimated depth. To maintain real-time performance (~29.32ms per frame), a sliding window expansion strategy merges new mesh segments into a global map without exhaustive re-processing.
Experimental Results
The system was validated using the FLSea real-world dataset and a modified ICL-NUIM synthetic dataset (underwater-ized using McGlamery’s light propagation model).
- Accuracy: In synthetic tests, the Absolute Trajectory Error (ATE) was reduced by more than half, from 0.055m to 0.024m, after applying image enhancement.
- Efficiency: The mesh representation is remarkably lightweight, using only 17,538 points to represent a scene that would require over 22 million points in a standard point cloud to achieve comparable visual quality.
- Interpretability: Qualitative results demonstrate that text on underwater markers, which is often illegible in voxel-based maps, becomes clearly readable in the proposed photorealistic mesh maps.
Conclusion
By combining a transformer-based enhancement front end with an efficient mesh-based back end, this system enables AUVs to perform high-fidelity 3D mapping in real-time. Future work will explore using the mesh data itself for loop closure detection to further refine localization accuracy.
Summary updated