CANSURF: An ASV-View Can Dataset for Detection and Tracking of Surface-Level Debris

This repository contains the raw CANSURF dataset and the augmentation script used to produce the full training set described in the paper. The repository contains the latest version of the dataset. To view all the versions of the dataset, please visit the Zenodo archive of the dataset: https://doi.org/10.5281/zenodo.20100657

Paper: Z. Aljundi, A. Moosa, M. Elemam, and Z. F. Rahmatullah, "CANSURF: An ASV-View Can Dataset and Benchmark for Detection and Tracking of Surface-Level Debris," 2025 8th International Conference on Signal Processing and Information Security (ICSPIS), 2025. DOI: 10.1109/ICSPIS67605.2025.11318414

---

Repository Structure

CANSURF/
├── dataset/
│   ├── train/
│   │   ├── images/       # Raw training images (.jpg)
│   │   └── labels/       # YOLO-format annotation files (.txt)
│   ├── val/
│   │   ├── images/       # Validation images (.jpg)
│   │   └── labels/       # Validation annotation files (.txt)
│   └── data.yaml         # YOLO dataset config
├── augment.py            # Augmentation script
├── train.py
├── yolov11_benchmark.py
├── yolov11_benchmark_sahi.py
└── README.md

---

Dataset

The raw dataset contains 7,171 images of aluminum cans floating on water. Annotations are in YOLO format (bounding boxes).

For the full augmented dataset used in the paper, run the augmentation script as is below without any augmentation type modification.

---

Augmentation

The augmentation script (augment.py) applies 10 augmentation types to each training image, expanding the training set. The validation set is left unaugmented to preserve evaluation integrity.

Augmentation Types

Type	Description
bright	Random brightness increase (factor 1.2–2.0) to simulate high sunlight
dark	Random brightness reduction (factor 0.2–0.8) to simulate low-light or shadow
color	Random contrast and color saturation boost to simulate camera variation
noise	Gaussian noise with random mean and standard deviation to simulate sensor noise
blur	Gaussian blur with random radius (1–4) to simulate motion or focus issues
weather	Random fog effect via Albumentations to simulate adverse weather
compress	JPEG compression at low quality (10–30) to simulate degraded footage
mosaic	Pixelation via downscale/upscale to simulate low-resolution conditions
hflip	Horizontal flip with corrected bounding box coordinates
vflip	Vertical flip with corrected bounding box coordinates

Flip augmentations correctly adjust YOLO bounding box coordinates. All other augmentations copy the original label file unchanged, as they do not affect object geometry.

Usage

Install dependencies:

pip install pillow numpy albumentations opencv-python

python augment.py

The augmented dataset will be written to ./augmentation_data/.

Customising Augmentations

To change which augmentations are applied, edit the augmentation_types list in the main() function:

augmentation_types = [
    'bright', 'dark', 'color', 'noise', 'blur',
    'weather', 'compress', 'mosaic', 'hflip', 'vflip'
]

Remove any types you don't want. Each entry in this list is applied once to every training image.

It is recommended to correspondigly remove any unwanted augmentation types in the augmentation_functions dictionary in augment_image_and_label():

augmentation_functions = {
    'bright': lambda i, l: adjust_brightness(i, l, bright=True),
    'dark':   lambda i, l: adjust_brightness(i, l, bright=False),
    'color':  adjust_color,
}

---

Benchmarking

YOLOv11 Benchmarking

Benchmarking scripts are provided to evaluate a trained YOLOv11 model on the CANSURF validation dataset using standard object detection metrics including:

• Precision
• Recall
• F1-score
• Mean IoU
• Confusion Matrix

The two scripts include a standard YOLOv11 benchmarking script as well as a YOLOv11 + SAHI benchmarking script. They perform inference on all images in the validation set and compare predictions against YOLO-format ground truth annotations.

Usage

Before running the benchmarking scripts, modify the following variables in the script:

MODEL_PATH = "path/to/your/trained_model.pt"
DATASET_PATH = "./CANSURF/val"

• MODEL_PATH should point to your trained YOLOv11 .pt weights file.
• DATASET_PATH should point to the validation dataset directory containing:
  • images/
  • labels/

Run Benchmark

python benchmark.py

Output

The script reports per-class metrics in the terminal and automatically saves a confusion matrix image:

yolov11_confusion_matrix.png

The confusion matrix is generated using predictions across IoU thresholds from 0.5 to 0.95.

---

Training

The script train.py is responsible for training the model based off of the CANSURF dataset. The script currently points to the YAML of the raw dataset, but should be changed if an augmented version is created. Refer to the Ultralytics documentation for further hyperparameter customisation.

Citation

If you use CANSURF in your research, please cite:

@INPROCEEDINGS{11318414,
  author={Aljundi, Zaid and Moosa, Abdullah and Elemam, Mostafa and Rahmatullah, Zahra F.},
  booktitle={2025 8th International Conference on Signal Processing and Information Security (ICSPIS)}, 
  title={CANSURF: An ASV-View Can Dataset and Benchmark for Detection and Tracking of Surface-Level Debris}, 
  year={2025},
  volume={},
  number={},
  pages={1-6},
  keywords={Weather;Wind;Aluminum;Pipelines;Detectors;Object detection;Benchmark testing;Water pollution;Reliability;Videos;Autonomous Surface Vehicles (ASV);Marine Debris Detection;Object Detection;Benchmark Datasets;Small Object Detection;Multi-Object Tracking;Domain Adaptation;Real-time systems},
  doi={10.1109/ICSPIS67605.2025.11318414}}