AI bird counting — demonstration reels
Two clips of the detector running on real footage. Every box is model output rendered frame by frame; nothing here is a mock-up. The measured numbers sit beside each clip, and so do the limits.
Barn counting — birds in the house
29.8mean birds per frame
48peak in frame
0.617mAP@50 (validation)
0.606recall
42.0sclip length
| Detector | YOLO11-s fine-tuned at 1280 px, 60 epochs, on 17 frames (647 boxes), held out on 4 |
|---|---|
| Inference | Sliced over 3x2 overlapping tiles, merged, then nested and impossible-size boxes suppressed |
| Hardware | Trained on a single RTX 3060; runs CPU-only on the platform |
Read this honestly. Training labels were generated automatically, so the mAP
figure measures agreement with the auto-labeller rather than with ground truth. This clip is a
floor-level view at working density, where birds are separable and boxes are the right tool. On a
wall-to-wall wide shot the detector fails outright — that case needs density estimation, not
detection. Source footage is public broiler-house video (CC0 / CC BY), not a customer site.
Dressed-bird classification — shackle line
Detection, tracking and grading on supplied line footage. Each box carries a persistent track ID, so a carcass is counted once however many frames it appears in. The grade is assigned against Appendix 001, and every result carries a confidence score — below threshold the system abstains and asks for a human.