Jeya Vardhan
← All workSoftware · Nov 2026

Airborne YOLO

A YOLO11 detector that tells aircraft, birds and drones apart in sky images, with a radar-style view and an honest look at where it fails.

Airborne YOLO, cover image
A radar scope: detections from the camera are drawn as blips by angle and distance, with per-class scores beside itCAMERA FRAME → POLAR → SCOPEDRONE 0.93BIRD 0.86AIRCRAFT 0.90CLASSMAP@50AIRCRAFT0.902BIRD0.864DRONE0.931YOLO11S · 1,778 VAL IMAGESOVERALL MAP@50 0.899
Fig. 1 — How it works

A camera watching the sky for drones is useless if it panics every time a bird flies past. This model sorts what it sees into three classes, aircraft, birds and drones, and draws the detections on a radar scope so you can tell at a glance where things are.

Results

Measured on the 1,778-image validation split at 640 px:

Class Precision Recall mAP@50 mAP@50-95
All 0.887 0.850 0.899 0.557
Aircraft 0.887 0.868 0.902 0.588
Bird 0.867 0.789 0.864 0.428
Drone 0.906 0.895 0.931 0.654

What I would trust less

  • Some validation images are near-copies of training images. 173 of the 1,778 (9.7%) share a source photo with a training image, so these scores are probably optimistic.
  • Birds are the weakest class. One tern gets labelled as an aircraft with 83% confidence.
  • The validation split also picked the checkpoint, and each setting was trained once, so there is no spread across seeds.

The radar view

Each detection’s position in the frame is converted to polar coordinates and drawn as a blip on a circular scope, which turns a busy image into something you can read in a second.

The guide

The repository doubles as an eleven-chapter guide to training YOLO on your own data: planning, labelling, splitting without leaking, training and using the model. Every example comes from this project, and 50 sample images let you try each step in minutes.