TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 16—18 · FACTUAL · EVOLVING
From photographs to a measurable model
Photogrammetry needs surveyed control to measure, and a splat gives a walkthrough rather than a surface.
When to use
Before any decision about an existing building: a measured-drawing assignment, a conservation report, an addition to an old house, a record of a structure about to be altered. The scan is the documentation step that every later heritage decision rests on.
The method
Two methods start from the same photo set and end in different places; decide which you need first. Photogrammetry gives a measurable model. Walk the building and photograph it in overlapping passes. Historic England's metric-survey specification asks for at least 60 per cent overlap between adjacent images and 40 per cent between strips, and its own drone briefs ask for 80 per cent front-to-back; plan for the higher figure. Keep a steady distance and even light; an overcast morning is better than noon, because hard shadows and uneven light break the matching between frames. Give the model control, or it has shape but no dependable size: a scale bar or a measured distance for a small subject; for a building, surveyed control points on the fabric — the specification asks for at least four per elevation, plus check points it does not use in the adjustment. Run the set through a photogrammetry tool. It finds the same points across photographs, works out where each camera stood, and builds a point cloud — a cloud of measured points — and from that a mesh you can section and measure. Gaussian splatting is the other method. From the same kind of photo set it builds a view you can move through in real time. It is a view-synthesis representation, made to look right from any angle; a measurable surface is not what it produces by default, and pulling one out of it is a separate step that loses accuracy. Then check the photogrammetric model against the building: measure three things on site with a tape — a door width, a column spacing, a plinth height — and compare; those are your check points. Write the error down. If the model is two per cent out, every dimension you take from it is two per cent out.
Watch for this
Trusting the surfaces the camera never saw. Photogrammetry invents nothing on purpose, but it smooths over gaps — the top of a cornice, the inside of a niche, a wall behind a tree — and the smoothed surface looks as solid as the measured one. Plain white walls, glass and polished stone also confuse it: the software needs texture to match points across frames.
Try it
Photograph one small built thing with strong texture — a temple plinth, a brick gateway, a carved stone post — in forty to sixty overlapping frames, with a tape in the scene. Build the model in any free photogrammetry tool. Measure three dimensions on the model and the same three on site. Report the difference as a percentage.
Prove it
Show one photogrammetric model beside its three on-site check measurements, with the error stated, and point to one surface in the model you know the camera never saw.
How it works
Photogrammetry is geometry, not machine learning. A method called structure from motion finds the same feature in many photographs and works out where the cameras were and where the points are; surveyed control gives the result its scale and its position, and check points measure its error. Machine learning comes in later. Gaussian splatting, published in 2023, represents a scene as a very large number of small soft blobs — 3D Gaussians — and adjusts them until views rendered from them match the photographs, which gives real-time walkthroughs at full HD. It is a rendering method first. A measurable surface has to be extracted from it, and that extraction is still a research problem: surfaces recovered from splats can be geometrically wrong where the photographs were ambiguous. Open projects such as Open Heritage 3D publish scan datasets of real monuments for anyone to download and section. EVOLVING because the methods change every year — what measures badly today may measure well by the next edition — and because phone-based capture is improving quickly. WORLD MODELS on this map is the frontier beyond this card; WHAT IT CAN TAKE IN covers how a photo set enters a model.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.