sapoepsilon/yolov11s-drone-detector

YOLOv11s Drone Detector

A YOLOv11s model fine-tuned to detect drones from RGB camera footage. Trained from the pathikg/drone-detection-dataset (54k images) and validated on the held-out Anti-UAV-RGBT test split (91 video sequences).

Held-out eval (Anti-UAV-RGBT, IoU=0.5)

metric	value
Precision	0.929
Recall	0.759
F1	0.836
Mean IoU on TPs	0.998
AP@0.5	0.741

Beats every other public single-class drone YOLO we found on HF on mAP@50 and mAP@50-95 (and is much smaller — 19MB).

Use

from ultralytics import YOLO
model = YOLO("sapoepsilon/yolov11s-drone-detector")
model.track("path/to/drone_video.mp4", tracker="bytetrack.yaml")

Training

base	ultralytics yolo11s.pt
dataset	pathikg/drone-detection-dataset (~54k images)
imgsz	640
batch	192 (3-GPU DDP)
epochs	25 (early-stopped from 30)
optimizer	AdamW, lr 4e-4, cosine
augmentation	mosaic + flip + close-mosaic at epoch 20
hardware	3× NVIDIA RTX 3090 (1× Ti)

Caveats

• Single class only (drone); doesn't distinguish drone subtypes
• Trained on YouTube-sourced RGB drone images; recall drops on long-range surveillance (Anti-UAV-RGBT shows ~76% recall on small/distant drones)
• Pair with a tracker (ByteTrack or BoT-SORT) for trajectory output

License

Apache 2.0.