Livestock stereo-camera sampling: internal lab note
A lab note on sampling bias when a stereo camera observes animals through one fixed scale.
- Owner
- lab-team
- Reviewed
- 2026-09-12 · Next review 2026-12-12
- Evidence level
- Internal test
- Citation scope
- Cite only as an internal lab note describing the sampling limitation; it is not a production or customer result.
Hypothesis
A low-cost stereo camera can estimate useful geometry only when the captured sample represents the full pen, not just animals that cross one scale.
Setup and hardware
Two CCTV cameras and a Raspberry Pi 5 calibrated as a stereo pair beside a single fixed scale; internal-approved observations only.
Two CCTV cameras · Raspberry Pi 5 · ESP32 sensors · YOLOv12 · custom stereo calibration
Method
Compare image-derived geometry with scale observations, then check whether the sampling path covers the whole pen before interpreting accuracy.
Result
The prototype demonstrates the capture path, but the single-scale sample is too small to support a publishable weight-accuracy claim.
Failed attempts
- Using the single scale as a proxy for the whole pen produced a biased evaluation set.
Limitations
- A fixed single scale samples too few animals to represent the whole pen.
- Image-to-weight accuracy is not publishable until the sampling problem is solved.
Next steps
- Design a sampling plan that covers pen zones and time windows.
- Repeat calibration and compare against independent scale observations.
Citation scope
Cite only as an internal lab note describing the sampling limitation; it is not a production or customer result.