IALA-10: A Benchmark for Fine-Grained Maritime Buoy Classification Aligned with the Maritime Buoyage System

Jan Lukas Augustin, Oliver Niggemann

Keywords: maritime computer vision, buoy classification, aids to navigation, IALA Maritime Buoyage System, foundation models, DINOv3, few-shot classification, benchmark dataset, safety-aware evaluation

TL;DR: We introduce IALA-10, the first fine-grained buoy classification dataset with 1,234 crops across 10 classes, and benchmark supervised and foundation models, finding DINOv3-B linear probing achieves 92.7% accuracy.

Abstract: Autonomous vessels must classify buoys by type under the IALA Maritime Buoyage System to comply with COLREGs, yet every major maritime detection dataset treats "buoy" as a single class. We introduce IALA-10, the first image dataset for fine-grained buoy classification: 1,234 labeled crops across 10 classes motivated by the IALA Maritime Buoyage System, produced via SAM3 zero-shot detection and human review from diverse web sources. We benchmark supervised fine-tuning, foundation model linear probing (DINOv2, DINOv3, CLIP), few-shot k-NN, and CLIP zero-shot baselines. DINOv3-B with logistic regression and domain-motivated band-order features achieves 92.7% accuracy and 0.894 macro-F1. A safety-aware cost matrix that penalizes navigationally dangerous confusions confirms the best model achieves zero cardinal north↔south confusion. CLIP zero-shot struggles on cardinal subtypes whose distinctions lie in vertical band ordering, a spatial property that text descriptions cannot capture well. Dataset, code, and benchmarks will be released upon acceptance.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8