AquaSH-Scenes: A Multi-Scene Underwater RGB–Laser Benchmark for Geometry–Color Mesh Reconstruction

We constructed multiple representative underwater scenes in a large-scale experimental water tank. These scenes cover the major structural categories commonly encountered in marine engineering and seabed exploration, including sandy seabeds, coral reefs, rocky reefs, seagrass-rich areas, benthic communities, and underwater pipeline structures. The sandy seabed scenes contain extensive low-texture regions, fine-scale sand ripples, and mild terrain undulations. They are designed to evaluate the ability of reconstruction methods to recover weakly textured and smooth surfaces. The coral reef scenes contain slender branches, highly curved edges, and complex occlusion relationships, enabling the evaluation of fine-structure reconstruction and the representation of irregular boundaries. The rocky reef scenes include rough surfaces, irregular protrusions, and locally shadowed regions, and are used to assess reconstruction performance on complex geometric relief. The seagrass and benthic-community scenes contain weak boundaries, partial occlusions, and regions with color blending, allowing the evaluation of robustness to underwater color degradation and structural discontinuities. The pipeline scenes contain regular engineered structures, cylindrical boundaries, and locally attached objects, and are intended to evaluate the preservation of geometric boundaries and color consistency for artificial targets. To simulate realistic underwater image degradation, data were collected under different illumination intensities and water-clarity conditions. Illumination was controlled using a lighting module mounted on the acquisition platform, with both high-light and low-light settings. In addition, a water-soluble organic dye was used to adjust wavelength-dependent attenuation in the water, thereby simulating underwater color casts and spectral absorption. These settings allow the dataset to cover typical underwater degradation factors, including low contrast, color shifts, and scattering-induced blur. Data Acquisition The data acquisition platform consisted of a remotely operated vehicle (ROV), an underwater 4K camera, an illumination module, and a high-precision laser scanning system. RGB images were captured using the underwater 4K camera mounted beneath the ROV, with an image resolution of 4096 × 3008 pixels. During acquisition, the camera maintained an approximately downward-facing orientation while the ROV moved smoothly along predefined trajectories. To ensure stable camera-pose estimation during Structure-from-Motion (SfM) processing and stable Gaussian optimization, an overlap ratio of approximately 30% was maintained between adjacent frames. For each scene, 70 consecutive images were collected, of which 55 were selected for training and 15 for testing. The test views cover different viewing directions, illumination conditions, and water-clarity levels. Geometric Ground Truth and Evaluation To obtain reliable geometric ground truth, each scene was scanned using a VOYIS Insight Micro-1000m Laser & 12MP Stills underwater laser and imaging system, producing high-precision laser point clouds. The laser point clouds were not used during model training and were used exclusively for quantitative geometric evaluation. For each scene, camera intrinsic and extrinsic parameters and an initial sparse point cloud were first recovered from the RGB image sequence using SfM. All evaluated methods were then initialized using the same camera poses and point-cloud conditions. Finally, the reconstructed meshes were rigidly registered to the ground-truth laser point clouds using the Iterative Closest Point (ICP) algorithm. Registration was restricted to valid seabed and structural regions, while water-medium regions, suspended-particle regions, and obvious non-structural noise were excluded.