RiSID: River Surface Image Dataset for Instance Segmentation of Floating Macroplastic Debris

TL;DR

This paper introduces a dataset of river surface images to help develop AI models for tracking plastic pollution in rivers.

Contribution

The paper introduces RiSID, a new annotated dataset for instance segmentation of floating macroplastic debris in river images.

Findings

RiSID contains 7,356 images with pixel-level annotations of floating plastic debris from 11 river sites in Japan.

The dataset is formatted in MS COCO and categorized into three different annotation sets for model evaluation.

RiSID supports the development of deep learning models for monitoring macroplastic transport in rivers.

Abstract

Rivers constitute a major pathway of macroplastic debris, which has the potential to have adverse impacts on marine ecosystems and oceans. It is essential to develop image-based technology for quantifying macroplastic debris floating on river surfaces and then grasping plastic transport from land. The river surface image dataset (RiSID) comprises 7,356 original images recorded at 11 sites on seven rivers during high-flow conditions in Japan, along with pixelwise segmentation annotations for floating macroplastic debris. The three annotation datasets were divided into seven, five, and two categories of floating anthropogenic debris to explore the model performance. The annotation data were packaged in a JSON file in the Microsoft Common Objects in Context (MS COCO) format, which is a common format for computer vision research on developing deep learning models. RiSID would be helpful for researchers to explore a better model for monitoring floating macroplastic debris via river surface images.

Figures