sailswarm-yolov8s-lars-seg

SailSwarm YOLOv8s-seg (LaRS typed instance masks)

YOLOv8s-seg trained on the LaRS benchmark for per-object instance segmentation of 8 maritime classes (boat_ship, row_boats, paddle_board, buoy, swimmer, animal, float, other) — the panoptic-style per-object typed masks used in the SailSwarm dashboard.

Results (LaRS val)

• Box mAP@50 0.44, Mask mAP@50 0.35. Per-class strong on boat (~0.56 mask) / paddle_board / swimmer; float (23 train instances) is dead; other weak. Boat-dominated; class-imbalanced.

Intended use & limits

• Undistorted RGB. LaRS-domain (pinhole/USV) — degrades on our fisheye (misses e.g. moored boats packed against a shoreline). Visualization-grade, not load-bearing; wants an on-domain fine-tune.

Provenance / license

Base: Ultralytics YOLOv8-seg (AGPL-3.0). Training data: LaRS (research/non-commercial — derived weights inherit its terms).

Citation

This model derives from the LaRS dataset/benchmark. If you use it, please cite:

@InProceedings{Zust2023LaRS,
  title={LaRS: A Diverse Panoptic Maritime Obstacle Detection Dataset and Benchmark},
  author={{\v{Z}}ust, Lojze and Per{\v{s}}, Janez and Kristan, Matej},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2023}
}