Synthetic Seabed Dataset

This dataset contains fully synthetic, automatically annotated images of marinedebris resting on the seabed. It was produced for a bachelor's thesis at theUniversity of Alicante that studies whether generative synthetic data can improveunderwater debris detection when real annotations are scarce. How it was generatedEach scene is built in two stages. A diffusion model (SDXL) first produces adebris-free seabed background. Every debris object is then inserted independentlyby inpainting it on a clean copy of the background with FLUX.1 Fill, and a singleSegment Anything (SAM) pass cuts the object into the scene through its mask andderives its annotation from that same mask. A final domain-randomised underwaterfilter (colour attenuation, turbidity and particles) is applied to the assembledscene. Because objects are generated from text prompts rather than cut from realimages, the dataset is leakage-free by construction: no real object instance canleak from a test set into training. Contents Images of synthetic underwater scenes with one or more debris objects. YOLO detection labels (class id + normalised bounding box) per image. YOLO segmentation labels (polygon from the SAM mask) per image. The exact folder layout is described in the accompanying documentation. Classes (19, following the CleanSea taxonomy)Plastic_Bag, Can, Packaging_Bag, Bottle, Fishing_Net, Plastic_Debris, Wood,Car_Bumper, Towel, Pipe, Rope, Glove, Basket, Squared_Can, Metal_Debris, Tire,Shoe, WashingMachine, Metal_Chain. Intended useThe data is intended for training and evaluating object detectors for underwatermarine litter. In the accompanying study it is used to train YOLOv8 detectors thatare always evaluated on the real CleanSea test set. It is meant to complement, notreplace, real underwater imagery. CodeThe generation pipeline and training/evaluation code are available in the linkedrepository. LicenseCreative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).