A project by TUM and Hochschule München.

Marine+

An improved model for detecting marine litter in coastal ocean pictures

Challenge

Finding garbage in the ocean: Can you beat the experts?

Researchers have annotated a dataset containing pictures from around the world to pick out pixels containing marine debris. Others have proposed a basic model to detect this litter. The challenge is: can you make their model better?

Team

Simon Chervenak, Oscar Röth, Mohamad Alkam,Zeynep Duran, Wasuwadee Kongdech, Ekaterina Gikalo

Contact: simonlcherv@gmail.com

About the prototype

The MARIDA dataset contains annotated images from twelve locations around the world. These images are annotated with many classes, including multiple types of water and plants, but our challenge focuses on marine litter. Our goal is to aid with the creation of a model that will help clean up our oceans and save our planet. The researchers made a basic neural network structure and provided us with tools to train and test that structure. To improve upon it, we changed the way it learns by encouraging it to focus on less common examples, since the marine litter was a very underrepresented portion of the images. We also shifted the locations of the data it was trained on to make it better at generalizing to unseen locations. Our solution increased accuracy by common AI metrics, but can still be improved by further examination of the data.

Outputs

  • pdf
    Project report

    Further information about the progress, milestones, and roadblocks.

  • Credits: Photo by Richard Carey/ Shutterstock