An open source dataset for Deep Learning-based visual detection of floating macroplastic litter

Abstract

<p>Plastic pollution of water bodies is a major environmental issue, as it can have harmful effects on marine life, riverine ecosystems and society as a whole. To mitigate the impacts of plastic pollution, accurate detection and quantification of macroplastic litter (plastic items > 5 mm) is of particular importance. In recent years, researchers and engineers have developed Deep Learning methods showing promising performances for detecting riverine macroplastic litter. However, there are several outstanding issues hindering the advancement of the field, including the lack of available data sources for training such models.</p> <p>Here, we present a new open source dataset for the detection of floating macroplastic litter. We generated the dataset from controlled experiments carried out in a small drainage canal on the TU Delft campus. The dataset features 626 different litter items including plastic bottles, bags and other plastic objects, as well as metal tins and paper litter. These items include household waste as well as litter recovered from canals in the Netherlands. We captured images with a resolution of 1080p and a linear field of view using two different action cameras and a phone, mounted on a bridge. The dataset consists of 10000 images, taken from two different heights (2.7 and 4.0 meters), two different inclinations (0 and 45 degrees from the horizontal), and two different weather conditions (sunny and cloudy sky).</p> <p>In this presentation, we provide information on the dataset and the experiments carried out to generate it. We also discuss the results of benchmark Deep Learning models for multi-class classification trained on the dataset, and their out-of-sample generalization ability to other case studies. While labels are currently available only for image classification, we aim to release annotations for object detection and image segmentation tasks in the future.</p>