Hold your thumb at arm's length and it covers the moon. That's the whole problem of measuring size from an image, in one gesture: a camera records angles, not sizes. A matchbox near the lens and a shipping container far from it can occupy exactly the same pixels, and nothing in the photo can tell them apart. Painters formalised this as perspective in the fifteenth century; computer vision formalised it as the projective camera model (the standard modern treatment is Hartley & Zisserman's Multiple View Geometry in Computer Vision); and every "measure anything with your phone!" app has to answer it somehow.
There are two honest ways out. One is to teach the phone its own optics — focal length, sensor geometry, lens distortion — and estimate distance, which is what AR measuring apps attempt, with accuracy that depends on hardware, lighting, texture and luck. The other is centuries older and needs no knowledge of the camera at all: put something of known size in the picture.
The trick: scale travels through the projection
Sentio's Dimension tool takes the second road. Photograph the object you care about with a reference beside it — a payment card or a sheet of A4 paper — then mark two line segments on the photo: one across the reference, one across the target. The tool computes one ratio and multiplies:
real size = (target pixels ÷ reference pixels) × reference's known size
Why is a payment card a legitimate ruler? Because it's one of the most tightly standardized objects you own. ISO/IEC 7810 fixes the ID-1 card format at 85.60 × 53.98 mm — every bank card, loyalty card and hotel key on earth, to sub-millimetre tolerance. A4 paper (297 × 210 mm, from the ISO 216 series) is the bigger sibling for bigger targets. You have carried precision references your whole life.
The geometry that makes the ratio valid is similar triangles: when two objects sit at the same distance from the lens, in the same plane, the projection scales both by the same factor — whatever that factor is. Focal length unknown? Cancels. Distance unknown? Cancels. Which phone, which lens, which zoom? All cancel. It's the platform's no-per-device-calibration principle in its purest form: the unknown camera appears in numerator and denominator, and divides out.
Where it's honest, and where it bends
Everything above holds exactly under stated conditions: reference and target in the same plane, at the same distance, viewed roughly straight-on. The tool's accuracy budget is the list of ways real photos violate them:
- Foreshortening. Tilt the plane and lengths compress along the tilt — a card photographed at a slant projects short, and everything measured against it reads long. Shooting square-on is the single highest-value habit.
- Different depths. A card on the table next to a box measures the tabletop's plane; the box's top edge lives centimetres closer to the lens and projects proportionally larger. Put the reference on top of, not beside, what you're measuring.
- Lens distortion. Wide phone lenses bow straight lines near the frame edges. Keeping both segments near the centre keeps the geometry honest.
- Tap precision. You mark the segments by hand, and a fingertip is a few pixels wide. The rule of thumb falls out directly: the bigger the reference in the frame, the tighter the answer. A card spanning a tenth of the photo makes those few pixels expensive; spanning half the photo, they're noise.
Do it well — coplanar, centred, reference large — and percent-level accuracy is realistic: a millimetre or two on a hand-sized object. That's the tool's ceiling, and it's stated rather than hidden. Notably, the saved measurement keeps both raw segments, the reference choice and the frame geometry, so the number is re-derivable and auditable later — a measurement you can re-check beats a number you have to trust.
Against the professionals
| Professional instrument | Phone Dimension tool | Where the gap bites |
|---|---|---|
| Vernier/digital calipers ($20–$200): ±0.02–0.05 mm, direct contact | Percent-level from a photo; no contact needed | Any tolerance work — shaft fits, thread sizing, 3D-printing calibration. Two orders of magnitude of precision are not negotiable |
| Tape measure ($5–$30): millimetre marks over metres, needs access to both ends | Measures from where you stand; works on things you can't touch or bridge | Long distances with nothing flat to lay a reference on; a tape is also simply faster when you can reach |
| Laser distance meter ($50–$500): ±1–2 mm over tens of metres, distance along the beam | In-plane sizes rather than distances; nothing to aim off a wall | Room dimensions, ceiling heights, site surveys — different geometry, different instrument, laser wins outright |
| Photogrammetry rigs / structured light scanners ($1,000+, calibrated targets and software) | One ratio from one photo; deliberately no camera model, no 3D | Curved surfaces, volumes, full 3D reconstruction — the tool measures flat-lay lengths and says so |
The role it actually plays: the Dimension tool is for sizes that would otherwise go unmeasured — the sofa candidate at the flea market, the alcove you'll be at the hardware store thinking about, the leaf you're tracking week to week, the fruit whose size feeds the ripeness features. Its competitor isn't the caliper; it's the guess. And unlike the guess, its result files under the object in your catalog with the photo, the segments and the date attached.
The Dimension tool shares the "clerical camera" family with counting and meter reading — see count, measure, read — and the wider question of when a phone should defer to a real instrument is the subject of the honest accounting.