Mapping the Catacombs: An Underwater Cave Segment of the Devil's Eye System

Michalis Chatzispyrou, Luke Horgan, Hyunkil Hwang, Harish Sathishchandra, Monika Roznere, Alberto Quattrini Li, Philippos Mordohai, Ioannis M. RekleitisarXiv 2025 · DOI · arXiv

Introduction

Underwater caves represent some of the most challenging environments for exploration and mapping. Critical for freshwater management and archaeological preservation, these systems are inherently dangerous and "GPS-denied." Traditionally, cave maps are produced manually by divers using knotted lines, compasses, and depth gauges—a slow, error-prone process that increases cognitive load in life-threatening conditions.

This research introduces a low-cost framework designed to automate the generation of accurate 3D maps using off-the-shelf hardware. By combining visual-inertial SLAM (SVIn2) with global structure-from-motion (SfM) optimization, the team successfully mapped the "Catacombs" segment of Florida’s Devil’s Eye cave system, producing both topological "stick maps" and dense photorealistic reconstructions.

Methodology and Hardware

The proposed pipeline utilizes a multi-stage approach to overcome the visual hostility of underwater environments, where standard sensors like LiDAR often fail due to light absorption.

The Hardware Rig

The team developed a custom sensor suite featuring three GoPro Hero9 cameras positioned to provide an overlapping field of view. This was paired with high-powered video lights and a Shearwater Perdix2 AI dive computer.

A diver maneuvers the custom camera and lighting rig through the narrow passages of the Devil's Eye cave system.

Software Pipeline

  1. Visual-Inertial SLAM (SVIn2): The framework estimates camera trajectories and generates sparse point clouds. A critical innovation is the integration of the dive computer’s pressure sensor data to provide an absolute depth reference, effectively correcting the vertical drift inherent in pure visual SLAM.
  2. Trajectory Alignment: Using an "AprilGrid" calibration target at the start of the dive, the independent trajectories from the three cameras are aligned into a single, unified coordinate frame.
  3. Dense Reconstruction: Keyframes and poses are processed through COLMAP and screened Poisson reconstruction to create smooth, high-fidelity 3D meshes.

Experimental Results: The Catacombs Dataset

The framework was tested in the "Catacombs" of the Devil’s Eye system at Ginnie Springs, Florida. This site was chosen for its complexity, featuring over 1km of interconnected branches and loops that require "side-mount" diving techniques due to narrow passages.

Trajectory and Metric Accuracy

The system successfully maintained consistency across a trajectory exceeding 1km. By anchoring the visual data to the dive computer's depth readings, the team achieved high metric accuracy in the vertical profile. The automated trajectory closely matched manual surveys conducted with specialized electronic mapping tools like the MNemo V2.

The generated topological map illustrating the complex, interconnected branches and loops of the cave system.

High-Fidelity Modeling

The dense reconstruction, utilizing over 40,000 keyframes, produced detailed models of six representative cave sections. These models allow for the scientific analysis of geological features, such as ceiling erosion and floor morphology, which were previously difficult to document with precision.

Discussion and Future Work

A primary insight from this work is the utility of standard dive computers as "ground truth" anchors for scale and vertical position in featureless underwater environments. However, significant challenges remain regarding computational bottlenecks; reconstructing a full 1km system in high density requires immense memory resources, suggesting a need for iterative or windowed reconstruction methods in the future.

Future iterations of this project aim to integrate 3D sonar to assist in low-visibility conditions and eventually port the entire pipeline onto Autonomous Underwater Vehicles (AUVs) for real-time, autonomous cave exploration.

Summary updated

BibTeX
@misc{chatzispyrou2025mapping,
  author = {Michalis Chatzispyrou and Luke Horgan and Hyunkil Hwang and Harish Sathishchandra and Monika Roznere and Alberto Quattrini Li and Philippos Mordohai and Ioannis M. Rekleitis},
  title = {{Mapping the Catacombs: An Underwater Cave Segment of the Devil's Eye System}},
  year = {2025},
  eprint = {2507.06397},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.06397},
}
Figure