Hlatar/Geoscan_pioneer_base_drone_detection_finetuned

CVproject — Drone Detection Weights

Fine-tuned YOLO model for drone detection. Used in the CVproject repository for stereo tracking and 3D trajectory reconstruction.

Training

• Base model: best.pt (continued training from previous doguilmak weights)
• Dataset: custom, single class — drone
• Epochs: 50 (training completed, early stopping did not trigger — patience=100)
• Image size: 640
• Batch: 8
• Optimizer: auto (selected by Ultralytics)
• LR0 / LRF: 0.01 / 0.01
• Momentum: 0.937
• Weight decay: 0.0005
• Warmup epochs: 3.0
• Amp: enabled (amp: true)
• Device: auto (effectively GPU; total training time ~1360 s)
• Seed: 0, deterministic: true

Metrics (validation, epoch 50)

Metric	Value
Precision (B)	0.9335
Recall (B)	0.9286
mAP@50 (B)	0.9640
mAP@50-95 (B)	0.5916

Losses on the final epoch:

Loss	Train	Val
box_loss	0.9916	1.4363
cls_loss	0.5046	0.6149
dfl_loss	1.0960	1.3429

Best values during training:

• mAP@50 ≈ 0.9664 (epoch 36)
• mAP@50-95 ≈ 0.5916 (epoch 50)
• Precision ≈ 0.9669 (epoch 33)
• Recall ≈ 0.9396 (epoch 36)

Training Curves

Loss and metric dynamics across epochs:

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("image.jpg", conf=0.25)

# Or on video / stream
results = model.predict(source=0, conf=0.25, stream=True)
for r in results:
    boxes = r.boxes
    for box in boxes:
        x1, y1, x2, y2 = box.xyxy[0].tolist()
        conf = float(box.conf[0])
        cls = int(box.cls[0])
        print(f"drone: {conf:.2f} @ ({x1:.0f},{y1:.0f},{x2:.0f},{y2:.0f})")