IALA-10: A Benchmark for Fine-Grained Maritime Buoy Classification

This repository contains the IALA-10 dataset and benchmark code for fine-grained buoy classification aligned with the IALA Maritime Buoyage System.

Paper: IALA-10: A Benchmark for Fine-Grained Maritime Buoy Classification Aligned with the Maritime Buoyage System Jan Lukas Augustin, Oliver Niggemann — CVPR 2026 Workshop on Maritime Computer Vision (MaCVi)

Dataset

1,234 labeled buoy image crops across 10 classes:

MBS Group	Class	Count
Cardinal	cardinal_north	37
Cardinal	cardinal_east	41
Cardinal	cardinal_south	25
Cardinal	cardinal_west	21
Lateral (A)	lateral_red	268
Lateral (A)	lateral_green	201
Safe Water	safe_water	96
Special	special	106
Non-MBS	mooring	84
Non-MBS	other	355

Images are organized in ImageFolder format under dataset/. Each image has a UUID filename. The file dataset/provenance.csv maps each UUID to its source category and source-image group identifier.

Sources: Wikimedia Commons, NOAA, Flickr, Geograph, Unsplash, the collision avoidance dataset (Gorczak et al., 2025), and additional manually collected images.

Quick Start

# Install dependencies
pip install torch torchvision scikit-learn transformers tqdm opencv-python pandas matplotlib

# Run full benchmark (extract features + evaluate)
python benchmark/run_all.py

# Extract features only
python benchmark/extract_features.py

# Evaluate only (requires extracted features)
python benchmark/evaluate.py

Reproducible Splits

• 5-fold CV: results/splits/fold_{0-4}.json — stratified group k-fold with source-image grouping
• Manifest: results/splits/manifest.json — full dataset index with paths and labels

Source-image grouping ensures all crops from the same photograph stay in the same fold, preventing data leakage from shared backgrounds.

Key Results

Method	Accuracy	Macro-F1	Safety Cost
DINOv3-B + Band + LogReg	0.927	0.894	0.137
DINOv3-B + LogReg	0.925	0.891	0.164
DINOv3-S + LogReg	0.916	0.866	0.198
ConvNeXt-Tiny (supervised)	0.882	0.786	0.282
CLIP-L/14 (linear probe)	0.823	0.659	0.464
CLIP (zero-shot, IALA)	0.399	0.287	1.127

Class Hierarchy

The 10-class taxonomy is defined in dataset/classes.yaml. See the paper for full details on the IALA Maritime Buoyage System alignment.

License

• Code: MIT License
• Dataset: CC BY 4.0

Citation

@inproceedings{augustin2026iala10,
  title={{IALA-10}: A Benchmark for Fine-Grained Maritime Buoy Classification Aligned with the Maritime Buoyage System},
  author={Augustin, Jan Lukas and Niggemann, Oliver},
  booktitle={CVPR Workshops (MaCVi)},
  year={2026}
}