Litter segmentation with LOTS dataset

Abstract

The marine ecosystem faces a significant threat due to the release of human waste into the sea. One of the most challenging issues is identifying and removing small particles that settle on the sand. These particles can be ingested by local fauna or cause harm to the marine ecosystem. Distinguishing these particles from natural materials like shells and stones is difficult, as they blend in with the surroundings. To address this problem, we utilized the Litter On The Sand (LOTS) dataset, which comprises images of clean, dirty, and wavy sand from three different beaches. We established an initial benchmark on this dataset by employing state-of-the-art Deep Learning segmentation techniques. The evaluated models included MultiResU-Net, Half MultiResU-Net, and Quarter MultiResU-Net. The results revealed that the Half MultiResU-Net model outperformed the others for most types of sand analyzed, providing valuable insights for future efforts in combating marine litter and preserving the health of our marine ecosystems.