Calibrated from a known length

Photo Caliper

Shop

Two photos side by side. Click two points across something whose size you know, type that size, and everything measured after that comes back in millimeters. Photos come from your own files: choose or drop one on each pane, and it stays cached in this browser until you forget it. Give measurements the same label in both photos (window, pair, hole span) and they line up in the comparison below. Ratios to a reference measurement need no calibration at all.

Session

3 · Compare A ↔ B

LabelA mmB mm B − AB / A A ÷ refB ÷ refratio Δortho mmA mm (known)B mm (known)

4 · Rectify: georeference one photo onto the other

Works like a GIS georeferencer. Switch to Tie points and click a feature in A, then the same feature in B: that is one numbered ground control point (GCP) pair. Controls are fitted, minimizing RMSE in the target photo's frame. Checks are held out: their error is the honest one. Targets are the unknowns (the rear hole, say) and are never fitted. Each direction is fitted on its own, not inverted from the other.

●control: fitted ■check: held out ◎target: transferred, never fitted ✕outlier (robust) ○excluded - -control hull lines: residual, GCP → where the fit puts it 0 → — line and ring color = error, same scale as heatmap
Detailed GCP table: coordinates, dX / dY, real mm, set
excl.#SetLabelRole sourcedest dXdYresidualleave-one-out real mm (x, y)

Click a row (here or in Point sets) to center both photos on that pair. Drag a marker in either photo to move it. "real mm" is optional: give 4 or more pairs their true position in mm (the corners of an 80 × 40 window, say) and each photo can be rectified to that plane on its own.

Transforms compared (★ lowest held-out error, not lowest fit)

Transformmin B → A fitLOOcheckscore A → B fitLOOcheckscore

Every transform here maps one flat plane to another, or bends to fit (polynomials, the thin-plate spline). The part is curved, so a point off the plane of the controls shifts, and no transform can undo that depth. Check points measure it. A spline passes exactly through every control, so its fit RMSE is always zero and means nothing: judge it by leave-one-out and checks. Keep targets inside the control hull; outside it, projective and higher polynomials extrapolate wildly.

Depth error beats every other error here. Calibrate on something the same distance from the camera as what you're measuring: anything FARTHER away than the reference reads short. Measured against the 30 mm washers the PR guard's hole span read 208.2, twice, from two photos, and it is really 220.7, confirmed by CAD and by a slicer. The whole 12 mm gap is one cause: the far hole sits ~45 mm lower than the reference, so it projects small. Use this for features lying in ONE plane at ONE depth: a hole offset, a boss diameter, a width across a flat pad. The moment your two points sit at different heights on a curved part, the number is short and you will not be able to tell by looking at it.

Consistency check. In two photos of the same part, a ratio that should be identical, such as window ÷ pair spacing, differing between them measures how different the camera geometry is. It bounds how much any other difference can be trusted: a B/A difference no bigger than that is the cameras, not the parts. Rectifying one photo onto the other removes only what the chosen transform can model; whatever the part's depth adds shows up as check-point error.

PR guard: what's known
Pair spacing30.0 ✓
Span, Ewan chord210
Span, Ewan arc220
Span, skid CAD220.88
Offset, Ewan8
Offset, skid CAD10.68
Offset, in the file9