Drive there. Set up the ladder. Hold the tape measure. Write numbers on a piece of paper. Drive back. Type numbers into Excel. Realise a measurement is missing. Drive back again. That's the reality of site surveys in the sign industry — unchanged for decades. I know this workflow from over 20 years of my own practice. A return trip for a forgotten measurement costs me on average 90 minutes: travel time, parking, re-measuring, driving back. At a rate of 65 euros per hour, that's nearly 100 euros per return trip — for a measurement I could have noted the first time. Three times a week, that's 300 euros lost or saved. With AI-powered photo recognition, there's a better way. One photo, one click, all measurements. Here I explain how it works technically and where the limits are. All features: Feature overview →

The basic principle

A photo of a storefront contains all geometric information — but in pixels, not metres. To get from pixels to real measurements, you need two things: an AI that detects and separates individual elements, and a reference measurement that enables the conversion.

The reference measurement is the key. If you know a standard door is 2.10m high, and the door in the photo is 400 pixels high, then 1 pixel = 5.25mm. From there you can calculate every other element in the photo. That sounds simple — the challenge is polygon segmentation: windows aren't rectangles, bay windows aren't squares, and a shop window with an arch has a completely different area than its bounding rectangle would suggest.

The multi-model cascade

PlotonIQ doesn't use one AI model but four — chained together. Each model has one job:

Claude Vision detects what's in the photo: windows, doors, signs, walls. It delivers bounding boxes — rough rectangles showing approximately where an object is. This is the first, coarse step.

Gemini Flash refines localisation. It understands text prompts like "all glass surfaces" and finds exact positions, correcting misclassifications from step 1.

Gemini 2.5 Flash analyses context: is this a shop window or a front door? Single glazing or double? Which materials are appropriate? This feeds into the material recommendation when you build the quote.

KI-Vision (Segment Anything Model) is the centrepiece. It produces pixel-accurate polygons — not rectangles. A window with an arch is captured with 30+ polygon points, not a rectangle that includes too much or too little area. That's the difference between "roughly right" and "good enough for a quote".

Practical example: storefront with 4 windows and a door

Last week I was standing outside a bakery. 4 shop windows, 1 entrance door with glass insert. Classic frosting project, frosted film with vinyl lettering for the business name.

Step 1: Photo with the iPhone, from about 3 metres, as frontal as possible. I was slightly off-angle because of a parked van — no problem, perspective correction handles that.

Step 2: Open in PlotonIQ. Tap a window. The AI instantly detects all 4 windows and the door as separate polygons. Duration: 8 seconds in this case.

Step 3: Set reference measurement. I tap the door and enter: height 2.10m. All other elements are automatically calculated. Result: Window 1: 1.42m × 1.18m = 1.68m². Windows 2 and 3 identical. Window 4 slightly smaller: 1.38m × 1.15m = 1.59m². Door: 0.92m × 2.10m = 1.93m² total glass area.

Step 4: Voice annotation. I dictate: "Windows 1-4: frosted film 3M Crystal. Door: opening hours white vinyl." The AI assigns materials to surfaces.

Result: An SVG sketch with all dimensions, areas, and material assignments. Total time: under 2 minutes. Traditional with tape and paper: 45 minutes to 1.5 hours including the return trip and typing up.

Accuracy: ±5%

The honest answer: AI measurement is not millimetre-precise like a laser distance meter. Accuracy is ±5%, depending on photo angle, image quality, and distance. A window that's actually 152cm wide will be recognised as 145-160cm.

For quotes, that's sufficient. You price with a material margin anyway — typically 10-15% extra for waste and errors. The AI inaccuracy falls within that buffer. For production — when you cut the film — you measure on site again. But the return trip to re-measure a forgotten dimension is completely eliminated because you already have all measurements in the system.

For comparison: manual measurement with a tape measure on complex facades also has 2-3% error — because people round numbers, sketches aren't to scale, or digits get transposed when typing up. The practical difference between manual and AI is smaller than you'd think.

Perspective correction

If you don't photograph exactly head-on (and nobody does, because cars, bike stands, or pedestrians are in the way), perspective distortion occurs. Elements at the edge of the photo appear narrower than in the centre. PlotonIQ corrects minor perspective errors automatically through homography transformation: the system detects parallel lines in the architecture and aligns the image accordingly.

For strong off-angle shots (more than 20° from the frontal axis) we recommend a second photo from a different position. The system can then combine both perspectives.

What doesn't work

  • Obstructed surfaces: If a tree, scaffolding, or van is in front of the window, the AI can't detect the obscured element. Use manual draw-and-snap to add it.
  • Night photos: Too little contrast between glass surface and frame for reliable segmentation. Photograph in daylight or with artificial light.
  • Highly reflective facades: Reflections from sun or opposite buildings confuse the segmentation. Overcast skies or diffuse light work best.
  • Fully glazed facades without visible frames: If glass meets glass seamlessly without a visible frame edge, the system can't separate individual surfaces.

In all these cases, use the manual draw-and-snap function — you draw the polygons yourself on the photo and the system calculates the areas from the reference measurement.

Frequently asked questions

Can I link the AI measurement directly to a quote?

Yes. The detected surfaces and material assignments flow directly into a new quote. The system automatically calculates material quantities and prices based on your master data. No manual transfer of numbers needed.

Does it only work for shop windows?

No. Facade signage, vehicle wraps (side view), signs on walls, roof surfaces of pagoda tents — anything you can photograph frontally. Vehicles have additional templates for common models on top of photo recognition.

How well does it work indoors?

As well as outdoors, as long as there's sufficient light. Even, artificial lighting — such as in a retail unit — often works better than direct sunlight that creates harsh shadows.

Do I need a special camera?

No. Any current smartphone is sufficient. An iPhone 12 or Samsung Galaxy A54 delivers adequate image quality. The angle and lighting matter more than the camera.

Conclusion

AI measurement doesn't replace the professional — it replaces the tape measure for initial surveys. Time saving: 80-90% per facade. No forgotten measurements, no illegible notes, no return trips. And most importantly: the manual fallback always works — you draw the polygons yourself when the AI struggles. Product design beats model hope: a system that offers a clean manual fallback for difficult cases is more valuable than one that claims to handle everything automatically.

Try the AI measurement on your next facade job. The first impression convinces faster than any explanation. Start for free →