Lab manual
lab1-pixel-size
Lab 1 · Pixel size calibration
- Goal
- Calibrate mm/px and verify against ground truth to within 0.05 mm.
- Estimated time
- 30min
- Level
- Introductory
- Prerequisites
- None. This is the first lab.
Goal
Calibrate pixel size on demo-2axis-vision and answer one question:
how far is your result from the truth?
Setup
- Machine:
demo-2axis-vision - Prerequisite calibration: run
C3 intrinsics and distortionfirst (do not skip it — step 5 explains why) - Time: about 30 minutes
Steps
File → Open machine → demo-2axis-vision, wait for all devices to connect.Calibration → Intrinsics and distortion, run with defaults. Note the reprojection error (should be < 0.15 px).Calibration → Pixel size. In prepare, set the acquisition span to 50 mm and 9 poses.- Record three numbers:
solved.mmPerPx,truth.mmPerPx,verify.deviationMm. - Change the span to 5 mm and the pose count to 25, run again. More poses, larger deviation — record it.
- Reset intrinsics (
Calibration → Reset → intrinsics only) and run pixel size directly. Note how much the residual moved compared with how much the deviation moved.
Pass criteria
| Item | Criterion |
|---|---|
| Deviation in step 4 | < 0.05 mm |
| Residual RMS in step 4 | < 0.15 px |
| Deviation in step 5 | clearly larger than step 4 (typically 3–5×) |
| Step 6 | residual grows a little, deviation grows more than 5× |
Common mistakes
- All poses crowded into the centre of the field. Beautiful residual, bad deviation. Span matters far more than pose count.
- Skipping distortion calibration. Step 6 makes you reproduce this on purpose.
- Unidirectional acquisition. Approaching every pose from the same direction hides backlash (Lab 3 covers it).
- Target parameters changed but not updated in configuration. A 0.1 % error in nominal spacing becomes a 0.1 % error in the result.
Review questions
- Going from 9 to 25 poses should improve precision by some factor in theory. Why was the result worse instead?
- In step 6 the residual RMS grew by less than 3× while the deviation grew by more than 5×. When, exactly, does a residual lie to you?
- On real hardware there is no truth. Design an experiment that estimates your deviation anyway. (Hint: gauge blocks, or two positionings a known distance apart.)
- A 0.1 % relative error in pixel size over 200 mm of travel is how many millimetres? Is that acceptable for your application?