New AI Trick Helps Tiny Cameras Spot Small Drones and Birds
This could put smarter drone-spotting eyes on cheap $80 hardware — no cloud needed.
Researchers built a training-time method that lets a compact YOLO11n student learn high-resolution spatial detail from a larger YOLO11m-P2 teacher — without changing the student's inference architecture. Under the Drone-vs-Bird validation protocol, YOLO11n-CSCWD reached 50.17% mAP@0.5 and 59.73% recall, beating the matched CA-YOLO11n baseline by 2.92 points in mAP@0.5 and 3.55 points in recall. On a Raspberry Pi 5 with NCNN-FP16 at 640x640, the 2.58-million-parameter student hit 50.32% mAP@0.5 at 82.32 ms latency — about 12.15 frames per second — with essentially the same runtime and memory as the baseline.
- A small AI learned from a much bigger AI during training, then ran alone — so it stays fast and cheap.
- Accuracy on spotting tiny flying objects rose about 3%, and it kept working on a brand-new dataset it had never seen.
- It runs at ~12 frames per second on a $80 Raspberry Pi 5, meaning smart cameras can work offline without the cloud.
Why It Matters
Cheaper, private, offline cameras could soon spot small drones and birds without sending your footage to the cloud.