SailSwarm YOLOv8n obstacle detector (Konstanz, semi-supervised)
YOLOv8n fine-tuned for the SailSwarm fisheye obstacle-detection module — the
Pi-4 deployment detector. Trained on hand-audited Day-1 Konstanz frames
(boat/person/structure) oversampled ×5, mixed with 450 down-weighted
GroundingDINO pseudo-labels.
Results (held-out eval clip, class-agnostic F1 @IoU0.5)
- F1 0.673 (P 0.68, R 0.66). Human-only baseline: F1 0.603. GroundingDINO
zero-shot oracle: F1 0.701.
Intended use & limits
- Undistorted fisheye frames; classes boat/person/structure (Day-1 footage has
no duck/buoy). Pi-4 deployable (export NCNN/ONNX/TFLite-int8).
Provenance / license
Base: Ultralytics YOLOv8 (AGPL-3.0 — applies to derived weights). Training
data: SailSwarm Konstanz captures + GroundingDINO pseudo-labels.
