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

About RiSID

The River Surface Image Dataset (RiSID) was developed to advance image-based technologies for quantifying macroplastic debris floating on river surfaces and to improve understanding of land-to-river plastic transport.
RiSID comprises 7,356 original images collected at 11 sites across seven rivers in Japan during high-flow conditions. Each image is accompanied by pixel-wise segmentation annotations of floating macroplastic debris.

To explore model performance under different levels of categorical granularity, we provide three versions of the annotation dataset, categorized into seven, five, and two classes. Annotation data are packaged in JSON format following the Microsoft Common Objects in Context (MS COCO) standard, which is widely used in computer vision research for deep learning model development.

We expect RiSID v1 to support researchers in developing and evaluating improved models for monitoring floating macroplastic debris in riverine environments.

This dataset was used in the study published in Frontiers in Earth Science: https://doi.org/10.3389/feart.2024.1427132.

Annotation Variants

Three annotation datasets are provided, each with a different level of category detail:

• 7-class labels: Drink bottles, Other bottles, Food containers, Shopping bags, Other bags, Other plastics, Non-plastics
• 5-class labels: Drink bottles, Food containers, Shopping bags, Other plastics, Non-plastics
• 2-class labels: Plastics, Non-plastics

These variants enable flexible use depending on the research objective and desired classification detail.

Repository Structure

The RiSID repository contains the following files:

• annotated.zip: 7,356 PNG images. Each file includes two panels:
  • Left: Original image
  • Right: Image with overlaid annotation information (segmentation masks, bounding boxes) using annotations_7cat.json
• images.zip: 7,356 cropped PNG tile images (1024 × 1024 pixels), extracted from the original frames
• annotations_7cat.json: MS COCO format annotations with seven categories
• annotations_5cat.json: MS COCO format annotations with five categories
• annotations_2cat.json: MS COCO format annotations with two categories

Note: The original 301 raw video files from which the 7,356 images were extracted were not deposited in the Zenodo repository because their total size exceeds 19 GB, making it impractical to provide them through this platform.

How to Use RiSID

We provide a Python script, quick_start.py, to help users quickly begin using RiSID. This script loads JSON annotation files and visualizes images with their corresponding segmentation masks.

Follow the steps below after downloading RiSID:

1. Unzip images.zip.

2. Install required libraries (NumPy, OpenCV, pycocotools) by running the following command. If you already have them installed, you may skip this step:

   pip install -r requirement.txt
   

3. Run quick_start.py with the JSON annotation file:

   python quick_start.py annotations_7cat.json
   

   Note: A JSON file is mandatory when executing the script. The output will display paired panels:

   • Left: Original image
   • Right: Image with overlaid annotation information (segmentation masks, bounding boxes).

   Example outputs can be found in annotated.zip.

Optional Arguments

You can customize input/output directories and transparency values:

python quick_start.py annotations_7cat.json -i images -o annotated_7cat -a 0.3

• -i: Input directory containing original images (default: images)
• -o: Output directory for visualized images (default: annotated)
• -a: Transparency value, between 0 (fully transparent) and 1 (opaque) (default: 0.6)