SailSwarm YOLO, on-domain fine-tunes on the audited Konstanz frames
Ultralytics YOLOv8/YOLO11 detectors fine-tuned on the human-audited fisheye
frames of the SailSwarm obstacle-detection corpus (Lake Constance, undistorted
864×648 fisheye; classes boat, buoy, duck, other, person, structure, see
data.yaml). They supersede NexusDwin/sailswarm-yolov8n-konstanz (June 2026,
LaRS-era labels).
| weights | val (129 images, 1779 boxes) P / R / mAP50 / mAP50-95 | note |
|---|---|---|
weights/yolov8n_audited_best.pt |
0.440 / 0.415 / 0.454 / 0.268 | trained on every audited frame |
weights/yolov8s_audited_best.pt |
0.477 / 0.460 / 0.462 / 0.295 | (+ seeds s1, s2; 11n/11s siblings) |
weights/yolov8n_noholdout_best.pt |
0.704 / 0.383 / 0.428 / 0.254 (68 held-out images, 1009 boxes) | leak-free: the audited holdout recordings removed from training; this is the detector behind the typed-evidence channel of the fusion scorer (GBT + typed AP 0.927 vs 0.874 without, 8 of 8 recordings) |
onnx/yolov8n_audited_640.onnx / _864.onnx and the v8s pair are the exports the box
runs (deploy/shadow_mode.sh typed; YOLOv8n-640 = 1.5 fps on a Raspberry Pi 4, 2
threads). Night-person use: v8s as a confirm-gate rejects 96 % of motor-as-person
teacher false positives while keeping 82 % of persons (2026-09-02).
Training logs: logs_yolo_*.log, yolo_noholdout_train.log. Recipe:
scripts/gpu_finetune/train_yolo.py in the SailSwarm-ObstacleDetection repo.
- Downloads last month
- 102