Dressed Bird Quality Control

Upload a processing-line video

The video is decoded, each carcass is detected and tracked down the line, and every bird gets one grade against the Appendix 001 standard.

AI bird counting — live demonstration reels

Two clips of the detector running: birds counted in the house, and dressed birds graded on the shackle line. Model output rendered frame by frame, with the measured numbers beside them.

Open the counting demos →

Barn bird counting demo

How it works

1 Detect & count carcasses 2 Track each bird down the line 3 AI vision grades damage per bird 4 Report + annotated video

Graded against the Appendix 001 Dress Bird Quality Standard (A / B / C / Reject).

Recent inspections

Diagnostics (internal)

Grading model

Trained on Appendix 001 reference photos plus in-domain shackle-line crops (mostly clean Grade-A birds auto-labelled and pending human confirmation). Headline accuracy is inflated by those easy normals; the hard defect classes remain weak and rare classes (broken, ammonia, reject) have almost no examples. Advisory only -- review every bird.

212
training photos
71%
grade accuracy (5-fold CV)
0.96
calibration temperature

Measured recall per grade

GradeABCREJECT
recall 87% 32% 63% 75%

Measured recall per defect

Torn skinScratch skinHaematoma / bruisesBroken (patah)Broken bruises (patah lebam)Ammonia burnOthersReject condition
14% 0% 50% 0% 0% 0% 0% 25%

Data readiness (labelled examples per grade × defect)

Torn skinScratch skinHaematoma / bruisesBroken (patah)Broken bruises (patah lebam)Ammonia burnOthersReject condition
A 5 5 15 2 0 1 0 0
B 13 6 17 0 4 1 0 0
C 10 6 10 0 10 5 5 0
REJECT 0 0 0 0 0 0 0 4