CANSURF Dataset Boosts Marine Debris Detection 12x
A new dataset named CANSURF has been launched, consisting of approximately 7.3k raw images sourced from videos, which have been annotated with bounding boxes. This dataset has been augmented using ten different techniques, resulting in around 57k training and validation images that cover a variety of lighting conditions and water states. It focuses on surface-level marine debris, particularly small reflective items such as aluminum cans, which hinder autonomous clean-up efforts. Various detector and detector-tracker pipelines designed for surface applications were evaluated. Training YOLOv11 on CANSURF enhances performance by 12 times compared to standard datasets. Findings indicate that YOLOv11+ByteTrack provides the most reliable tracking with minimal identity switches, while YOLOv11+SAHI improves recall for distant cans, albeit with reduced precision in full-context scenarios. This research is detailed in arXiv:2605.16774.
Key facts
- CANSURF dataset comprises ~7.3k raw images from videos
- Dataset expanded to ~57k images via ten augmentation types
- Training YOLOv11 on CANSURF boosts performance 12x over generic datasets
- YOLOv11+ByteTrack yields most stable tracks with fewer identity switches
- YOLOv11+SAHI increases recall on far-field cans but lowers precision
- Dataset targets small reflective debris like aluminum cans
- Published as arXiv:2605.16774
- Benchmarked detector and detector-tracker pipelines for surface operations
Entities
Institutions
- arXiv