Underwater3D-36K

We present Underwater3D-36K, a synthetic dataset designed for large-scale training and evaluation of underwater 3D reconstruction. It contains 1,000 scenes, each with 36 multi-view underwater images, ground-truth depth maps, reference RGB images (GT), and 3D mesh models. The dataset is based on high-precision 3D mesh models from public sources like USGS and Google Earth Engine, representing exposed terrain regions with reliable geometry.Rendered using the Blender engine, the dataset incorporates optical parameters such as refractive index, attenuation, and scattering to simulate realistic underwater effects. Each scene includes 36 images from different viewpoints, enabling multi-view geometric learning. The dataset ensures accurate pixel-wise alignment between images and mesh models, supporting depth estimation and mesh generation tasks.Underwater3D-36K significantly expands on previous datasets, providing a comprehensive platform for training, evaluation, and benchmarking underwater 3D reconstruction methods.