Acquired by the Image Processing and Pattern Recognition group as part of its research on underwater 3D scene reconstruction, the Mermaid Underwater Dataset is a unique open-access resource designed to support research in underwater photogrammetry, 3D reconstruction, and visual navigation. The dataset consists of underwater images acquired at a depth of around 19 meters at the “La Sirène” site (Lion-de-Mer, Saint-Raphaël, France) during the 2022 Submeeting workshop. The surveyed area covers approximately 150 m² and includes a statue of a mermaid, a sandy seabed with fragments of Roman tiles, and a rocky area with marine life distributed across the scene. The dataset was captured by divers using a single camera under natural lighting conditions.

The dataset consists of 1,250 high-resolution RGB images (3840×2800 px), totaling about 15 GB. The ground sampling distance (GSD) ranges from 0.5 mm to 4 mm, allowing for sub-millimetric detail. Along with the images, the dataset includes a dense set of supporting metadata: a calibrated micro-geodesic reference network (65 m², 5 ground control points), full camera calibration data, precise camera trajectories, and pose information.

This dataset, available on the SEANOE platform, can support a wide range of tasks, including photogrammetry, 3D reconstruction, underwater visual SLAM, structure-from-motion (SfM), multi-view stereo (MVS), feature matching, 6DOF pose estimation, relocalization, loop closure detection, trajectory analysis, color correction, object detection, and more.

From left to right: a team member during a dive in the scene, a close-up on the mermaid with extracted keypoints, trajectory (successive camera poses) of the images in the dataset, and obtained 3D reconstruction