Uncovering Anomalous Events for Marine Environmental
Monitoring via Visual Anomaly Detection
Laura Weihl, Stefan Hein Bengtson, Nejc Novak, and Malte Pedersen
Joint Workshop on Marine Vision @ ICCV’25
The first publicly available multi-annotator benchmark dataset for visual anomaly detection in underwater scenes.
This dataset contains 25 underwater videos from two marine locations in Denmark with annotated anomalous events (fish, crabs, and biological activity). Each video was annotated by 16 people, providing soft labels that capture annotation uncertainty and consensus event boundaries for temporal evaluation.
Key Features:
- 25 videos:
- 10 from Scene A – Hundested Harbour, Denmark
- 15 from Scene B – The Limfjords-bridge in Aalborg, Denmark
- 15,083 total frames with frame-level soft labels
- 16 annotators per video capturing subjective nature of “interesting” events
- Multi-annotator consensus labels for event boundaries
- Two distinct underwater scenes with different visual characteristics
The dataset supports research in visual anomaly detection, particularly for applications in marine environmental monitoring.
If you use the dataset or the code for your work, please cite:
@inproceedings{weihl2025uncovering,
title={Uncovering Anomalous Events for Marine Environmental Monitoring via Visual Anomaly Detection},
author={Weihl, Laura and Bengtson, Stefan Hein and Novak, Nejc and Pedersen, Malte},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025},
pages={2085--2094},
year={2025}
}