How Photogrammetry Works

Start with the thing your own visual system does for free. You have two eyes a fixed distance apart - the *baseline* - and each sees the world from its own viewpoint. Anything close shifts a lot between the two views; anything far shifts only a little. Your brain reads that shift, called parallax, and turns it into a sense of distance. Photogrammetry rebuilds exactly this reasoning from photographs, except that instead of two eyes it can use dozens or hundreds of camera positions, and instead of a brain it uses geometry.

Here is the geometry, stripped bare. A camera, crudely, projects the 3D world onto a flat 2D image: every point in front of the lens lands somewhere on the sensor, and a straight line - a *ray* - runs from that point, through the lens, to its spot in the image. A single photo is therefore ambiguous about depth: a point near and a point far can land on the very same pixel, so one image alone cannot tell you how far away anything is. But photograph the same point from a second position and you get a second ray. Two rays to the same real point, from two known camera positions, intersect at exactly one place in space - and that intersection is the point's 3D position. This is triangulation, and it is the heart of the method.

Notice what the software must know to make that work: *where each camera was and which way it pointed* (its pose), and *which pixel in one photo is the same real point as which pixel in another* (the correspondence). Remarkably, photogrammetry software solves for both at once from the photos themselves - we will unpack how in Lesson 2.3. For now hold the intuition: every surface point you want to reconstruct has to be visible, and clearly identifiable, in at least two, and preferably many, overlapping photographs taken from different angles. The more views of a point, and the better spread their angles, the more rays intersect on it and the more confidently its position is fixed. That single sentence explains almost every capture rule that follows - why overlap matters, why you orbit a subject, why blank walls and mirrors defeat the method. Depth is not in any one photo; it lives in the parallax *between* photos.

Two photos fix one point. Photogrammetry scales that up in two directions at once - more points and more photos - to build a complete model. Modern software begins by finding thousands of distinctive little features in each photo (a corner of a window, a crack in plaster, a speck on stone), matching the same feature across overlapping images, and using those matches to solve simultaneously for all the camera poses and a first, sparse cloud of 3D points. Then it goes dense: knowing exactly where every camera was, it can work across the overlapping images pixel by pixel to estimate a 3D position for a huge proportion of the visible surface, producing a dense point cloud of millions of points. Skin a continuous surface over that cloud and you have a mesh; drape the original photographs back over the mesh and you get a photorealistic textured model. Point cloud, mesh, texture - the same deliverables laser scanning produces, arrived at from pictures.

The practical consequence is that photogrammetry is hungry for coverage and overlap. Every part of the subject you care about must appear in several photos, seen from several angles, with neighbouring photos sharing a large fraction of their view so the software can chain them together. A lone photo of the back of a building contributes nothing reconstructable on its own; a ring of overlapping photos all the way around it reconstructs the whole thing. This is why a photogrammetry 'capture' is not a few arty shots but a *systematic sweep* - orbits, grids, rows - designed to leave no gap and no surface seen from only one direction.

It is also why photogrammetry rewards resolution and punishes haste. Sharper, higher-resolution photos give more and finer features to match, so detail and accuracy improve; blurry, dim or low-overlap photos give the software too little to work with, and the reconstruction degrades into noise, holes or distortion - or simply fails to build. Understanding the model as 'rays crossing, densified over a well-covered surface' tells you immediately what good input looks like: many sharp, overlapping, well-lit photos of a textured subject, taken from a deliberate spread of positions. Get the input right and the 3D almost makes itself; get it wrong and no amount of processing will rescue it.

Here is a limit that surprises people, and it matters enormously for anyone who intends to *measure* from a photogrammetry model. The geometry of intersecting rays recovers the shape of a scene and the *relative* positions of everything in it beautifully - but, from photographs alone, it does not know the absolute scale. A set of photos of a real doorway and a set of photos of a perfect dolls-house model of that doorway can produce the identical reconstruction; nothing in the pictures themselves says which is two metres tall and which is twenty centimetres. Photogrammetry gives you a faithful model that could be any size. To make it a *measurement*, you must tie it to something of known real-world dimension.

The everyday fix is a scale bar or a known reference: a ruler, a printed target of a precisely known length, a calibrated bar, or any object whose true dimension you have measured by other means, placed in the scene and captured in the photos. Tell the software 'these two points are exactly 1.000 m apart' and it scales the entire model to match - now every dimension you take off it is in real units. For better results you use several scale references spread through the scene, not one. For work that must sit in real-world coordinates - georeferenced to a site, a map or a survey datum - you go further and use surveyed control points whose true coordinates a surveyor has established, which also let you detect and reduce distortion across a large capture (the drone case in Lesson 2.4 leans heavily on this).

And this is precisely where the professional boundary sits. Putting a ruler in frame gives you a usefully scaled model for design and as-built work. But the *stated, checkable accuracy* of a deliverable, survey-grade scale, and anything georeferenced or legally binding depend on proper control, verified equipment specification and method - and belong to a licensed surveyor or geospatial professional, not to an app's confidence number. Treat every accuracy, resolution or range figure in this module as illustrative of a principle, never as a specification. The honest habit: always introduce known scale, always verify a few real dimensions against a tape on site, and know when the job has crossed into survey territory that needs a professional.

Photogrammetry's great virtue is accessibility. The instrument is a camera - a DSLR, a mirrorless body, a drone, or the phone in your pocket - so the cost of entry is a fraction of a survey-grade laser scanner, and the technique scales from a single ornament to a building facade to an entire landscape just by changing the camera platform. Because the raw data *is* photographs, the output carries rich, true colour and photographic texture: a photogrammetric model of a carved stone screen or a weathered wall looks like the thing, which makes the method a favourite for heritage documentation, facades, visualisation and any job where appearance matters as much as metrics. For much of Indian practice - cost-sensitive, heritage-rich, often without access to high-end scanners - this accessibility is decisive.

The weaknesses are the mirror image of how it works. Because it lives on matching features, photogrammetry needs texture and even light. Blank, flat, uniform surfaces (a plain painted wall, a clear sky) give nothing to match; shiny, transparent or reflective surfaces (glass, polished metal, water, mirrors) lie to it, because what they show changes with viewpoint; and moving things, deep shadow, harsh glare or blur wreck the matching. It is also computationally heavy and slow: turning hundreds of photos into a dense model can take hours of processing on a capable machine, unlike a laser scanner that delivers points almost live. And as we have seen, scale and absolute accuracy demand deliberate care - references, control and verification - in a way that a calibrated scanner handles more directly.

So where does it fit in the reality-capture family? Reach for photogrammetry when the subject is well-textured and well-lit, when budget or access rules out a scanner, when colour and texture are the point, and when you can invest capture effort and processing time - objects, facades, heritage detail, sites from the air. Reach for (or add) laser scanning when you need fast, direct metric accuracy, when surfaces are difficult or poorly lit, or when the job is large and complex - the subject of Module 3. In practice the two are often *combined*: a laser scan for a trustworthy metric skeleton, photogrammetry or the neural methods of Module 4 for rich visual detail. Knowing how photogrammetry works is what lets you make that choice deliberately rather than by default.