Amphitrite v1.0: An underwater database for marine debris and biodiversity

Overview:

The database (named Amphitrite v1.0) is a raw, unaugmented set which has been labelled to 16 categories with quality checked, bounding boxes, suitable for object recognition. For technical validation, we trained Amphitrite v1.0 on the latest versions of YOLO which scored a mAP50 of 0.42.

Background:

There is a plethora of promising research suggesting the use of computer vision for marine debris tracking and detection, however, most research concludes that the lack of data cannot provide conclusive results.

There are similar datasets available that are open source, however to our knowledge, collections of both synthetic debris and marine life options are limited and most sets do not cover a wide enough array of classifications with a diverse range of conditions to produce more robust conclusions. Collection of underwater debris footage is particularly challenging to accumulate, due to limitations such as size, quality or diversity; so this set is a collaborative and collective effort of marine organisations, charities and divers.

Dataset Details:

The dataset consists of 1,583 images that are observations collected around the world of underwater marine debris, fauna and flora which have been labelled (with bounding boxes). The images are in varying depths, locations and visibility to create as much of a robust and close representation of a marine environment as we can. There are 5,241 instances across 16 different label categories.10 labels fall under marine debris: Plastic Wrapper, Plastic Object, Plastic Container, Metal Container, Metal Object, Glass, Wood, Cloth, Rubber and Net. Biodiversity includes the remaining 6 classes: Fish, Coral, Jellyfish, Crustacean, Seaweed and Shell. We excluded items smaller than macro debris as microplastic require different methods of data collection. We also excluded humans, mega fauna and background items (rocks, sand, ROV etc).

File Structure:

The ‘dataset’ folder contains the entire set so that the user can split it into a train/val/test split in the way they wish to. Within the folder, there’s an ‘images’ folder (all .JPEG format) and a ‘labels’ folder. The ‘labels’ folder contains a folder of txt files, ready for use with YOLO and a ‘xml’ folder with the original labels in the xml format, allowing the user both options. A README.txt file describes in detail the photographs contributed and who they were donated by, down to file names, to ensure that all photographers are properly credited for their images.

Additionally, the dataset folder includes the Labelling Key document for anyone who wishes to use it and continue to add their own data. As well as the dataset.yaml file (which includes the current 16 classes) for ease of use when training new models.

Data Collection:

The dataset is held under a CC BY NC SA 4.0 license, so that researchers and students are free to use the data but it is not available for commercial use. The reason for this is because plenty of organisations have collected this data and still own the intellectual property - many have expressly indicated that they do not wish for the data to be commercial. This collection would not be possible without the hard work of the many organisations who collect underwater footage and donated their work to us:

• Scuba Divers: Jack Alexander, Phil Robinson, Maria Shokouros-Oskarsson, Zoe Moorton
• Underwater Photographers: Bo Pardeau
• J-EDI and JAMSTEC
• Nic Emery at The Honest Diver
• Christy Johnson at Nudi Wear, Hawaii
• Bryce Groark at Living Ocean Productions
• Margot Watson at Kraken Diving
• Megan Lamson at Hawai’i Wildlife Fund
• Cheryl King at SHARKastics
• Steve Laing at Underwater Adventures
• Renee Street at Bali Aqua Nusa Penida
• Minna Zakheou at Viking Divers, Cyprus
• Marine Conservation Society, UK
• Dr Jerome Harley at University of Seychelles, Seychelles
• Dr Melanie Bergmann at The Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research, Germany
• Underwater Photographers on Pexels

To the ocean, From thalassophilia all around the world.