UAVid Fine-tuned Road Segmentation Model

U-Net road segmentation checkpoint (base_channels=32), pretrained on
DeepGlobe and fine-tuned on binary road masks derived from UAVid.

Test metrics

Metric	Score
IoU	0.7168
Dice	0.8350
Precision	0.8149
Recall	0.8561
Pixel accuracy	0.9550
BCE + Dice loss	0.1963

Best checkpoint: epoch 47
Validation IoU: 0.6998
Validation Dice: 0.8234

Model

• Architecture: four-level U-Net
• Base channels: 32
• Input size: 256×256 RGB
• Prediction threshold: 0.5
• Fine-tuning dataset split: 536 train / 67 validation / 67 test

Checkpoint

best_model_uavid_finetuned_iou7168.pth

SHA-256:
32fd1735b805f27bf31048f14a276ff3c3d5cba78b7ca2f63c6d49ea6dea1468