polylane-roadway-yolo11l-seg
YOLO11-l-seg fine-tuned for roadway (asphalt road surface) segmentation in aerial drone footage of urban / residential streets.
Training
- Base model:
yolo11l-seg.pt(Ultralytics, COCO-pretrained) - Data: ~2 973 frames (2 646 train / 327 val) automatically annotated with SAM 3 text-prompted segmentation across 6 drone videos (~50 K frames total), Kashkadarya region (Uzbekistan)
- Split: train = 5 videos, val = 1 video (no overlap)
- Image size: 1280
- Batch size: 16 × 3 GPUs (NVIDIA B200)
- Epochs: 50, AdamW, cos LR, lr0=0.001, mixup=0.1, copy_paste=0.1
- Wall time: ~28 min on 3× B200
Metrics (val on held-out video)
| Metric | Best epoch (38) | Last epoch |
|---|---|---|
| Box mAP50 | 0.73 | 0.73 |
| Box mAP50-95 | 0.44 | 0.42 |
| Mask mAP50 | 0.604 | 0.58 |
| Mask mAP50-95 | 0.415 | 0.40 |
The roadway is a "stuff" class (one large region per image), so instance mAP under-reports semantic-segmentation quality. Visual mask IoU is closer to 0.70–0.80 on most frames.
Limitations
- Trained on residential / village aerial drone footage only — generalisation to highway/dashcam not validated.
- Training labels come from SAM 3 auto-annotation, which fails (~5–15 %) on far frames where grey rooftops are confused with asphalt. Some noise inherited.
Usage
from ultralytics import YOLO
model = YOLO("best.pt")
results = model.predict("frame.jpg", imgsz=1280, conf=0.25)
results[0].show()
License
Apache 2.0