In this guide
How Photogrammetry Turns Photos into a 3D Model
A phone camera, a turntable and a laptop can digitise a small object without any laser hardware at all. Photogrammetry is the technique that makes that possible: it rebuilds a three-dimensional shape from ordinary two-dimensional photographs, by finding the same surface details in many views and working out where those details sit in space. This guide explains the pipeline step by step, the capture rules that decide whether a scan succeeds, and the failure modes that leave a mesh full of holes.
Why this matters now
On 29 October 2026, ELEGOO launched the M1 Ultra, its first all-in-one handheld 3D scanner, combining blue-laser scanning modes with infrared structured light and a reported laser accuracy of up to 0.02 mm. Commercial scanners like it are impressive, and far beyond a student budget. Photogrammetry is the student-scale alternative to the same goal: instead of projecting laser lines or light patterns onto an object and measuring their distortion, it squeezes the geometry out of photographs the student already knows how to take. The trade is honest and worth stating plainly: photogrammetry costs almost nothing in hardware, but it demands careful capture and patient processing, and it cannot match a commercial scanner's accuracy or handle every surface.
A companion student build puts this guide into practice: a bench-top turntable scanner that photographs an object at 36 known angles and rebuilds it into a printable mesh. See the build here: https://projectech.in/projects/m1-ultra-inspired-turntable-3d-scanner-photogrammetry-esp32/
How photogrammetry works
Step 1: Find the same detail in many photos
Every photograph of an object contains hundreds of small, distinctive details: a corner, a speckle, a scratch, an edge where light changes sharply. Photogrammetry software scans each photo for such features and describes each one numerically, so the same physical detail can be recognised again in another photo even when the viewpoint, lighting and scale have changed slightly. This matching is the foundation of everything that follows. If two photos share no recognisable details — because the object is featureless, shiny, or photographed from too far apart — there is nothing to match, and the reconstruction stalls at this step.
Step 2: Work out where each camera was
Once features are matched across photos, the software solves a geometry puzzle. A single matched detail seen in two photos defines two sight lines, one from each camera position; where those lines cross is the detail's position in space. That is triangulation, the same principle surveyors have used for centuries, applied thousands of times over. At the same time, the software estimates the camera positions themselves, since the student did not measure them: it starts from rough guesses and refines them until the projected positions of all matched details line up with where they actually appear in every photo. This joint refinement is called bundle adjustment, and it is why a photo set with plenty of overlap converges to a consistent model while a sparse set drifts and warps.
Step 3: From a sparse cloud to a dense cloud
The output of bundle adjustment is a sparse point cloud: a few thousand points marking the matched details, plus the recovered camera positions. That already describes the object's shape in outline, but it is far too thin to print. The next stage, dense reconstruction, compares small patches of pixels between neighbouring photos and estimates a depth for nearly every pixel, filling in the surfaces between the matched details. The result is a dense cloud of millions of points that follows the object's surface closely, including gentle curves that no single feature marked.
Step 4: Mesh, clean, scale, export
A point cloud is not yet a printable object. Meshing connects neighbouring points into triangles, producing a continuous surface. Cleaning removes stray points, fills small holes, and trims the background and turntable plate out of the model. Then comes the step students most often skip: scale. Photogrammetry recovers shape but not absolute size, because a small object photographed close up and a large object photographed far away can produce identical images. The fix is a reference of known size in the scene — for example, a 20 mm calibration cube scanned in the same session — which lets the builder scale the mesh to real millimetres and measure the error with a vernier caliper. Only then is the mesh exported, usually as an STL file, for 3D printing or measurement.
Capture rules that decide success or failure
Overlap beats resolution
Two neighbouring photos should share roughly 60 to 80 percent of their content. That is why a turntable stepping 10 degrees at a time works well: each step moves the view only a little, so every surface detail appears in many consecutive photos and the matching stage has plenty to work with. A phone with a modest sensor and good overlap will beat an expensive camera with wide, disconnected jumps every time.
Fix everything except the angle
The camera must not move, the focus must not change, and the framing must stay identical across the whole set. On a turntable rig the object rotates while the camera stays still, which gives the same effect as walking around the object with a fixed camera. Changing focus mid-session changes the effective lens geometry and quietly corrupts the camera calibration the software is estimating.
Light evenly, and keep it still
Soft, even, constant lighting is the single biggest quality lever after overlap. Harsh shadows move across the surface as the object turns and create false features; mixed daylight changes colour and brightness during a session; flickering room lights beat against the camera shutter. A ring light around the object, with room lights kept steady, removes most of these problems at negligible cost.
Let the turntable settle
A stepper motor does not stop dead: it overshoots and vibrates for a fraction of a second after each step. A photo taken during that vibration is blurred, and a blurred photo poisons feature matching not only for itself but for its neighbours. A short settle pause — two seconds is a practical rule — after every step costs just over a minute per revolution and protects the whole reconstruction.
Tools students actually use
Meshroom and COLMAP are the two open-source tools most student builds rely on. Meshroom packages the full pipeline behind a node-based interface and is the friendlier first choice, though its complete pipeline benefits from a capable graphics card. COLMAP is more modular and widely used in research, and it runs well on a CPU for smaller photo sets. Both are free, both accept ordinary phone photos, and both produce meshes that can be cleaned in free tools and exported as STL. Neither tool rescues a bad photo set: if the capture rules above were broken, the honest move is to re-shoot, not to tune settings until something plausible appears.
From mesh to printed part
A raw mesh is rarely print-ready. Typical cleanup includes removing the turntable surface, closing holes on the underside the camera never saw, reducing triangle count for a manageable file size, and checking the result is watertight — a closed surface with no gaps — before slicing. The underside deserves a note: a turntable scan never sees the face the object rests on, so that face is always reconstructed by hole-filling or left flat. For display models that is fine; for parts that must mate with something else, it is a real limitation to design around.
Common failures and what causes them
| Symptom | Usual cause | Fix |
|---|---|---|
| Photos refuse to align | Too little overlap, or a featureless surface | Smaller steps, more overlap, add matte texture |
| Mesh full of holes | Blurred photos, moving shadows, glossy surface | Settle pause, steady lighting, matte coat |
| Warped or bent model | Camera moved or refocused mid-session | Fixed tripod, locked focus, re-shoot |
| Wrong size in print | No scale reference in the scene | Scan a calibration cube of known size |
| Noisy, lumpy surface | Low light forcing high sensor noise | Brighter even lighting, lower sensor gain |
Frequently asked questions
Do I need a laser or a depth camera to scan in 3D?
No. Photogrammetry needs only overlapping photographs and software; lasers and structured light are faster and more accurate on difficult surfaces, but they are a different method, not a requirement. The student turntable build linked above uses a phone camera alone.
How many photos does a good scan need?
For a small object on a turntable, 36 photos — one every 10 degrees — is a practical baseline that gives strong overlap all the way around. Fewer, wider-spaced photos save time but risk alignment failures; many more photos improve coverage of awkward shapes but lengthen processing, which can already exceed an hour for a 36-photo set on a modest laptop.
Why do shiny and transparent objects fail?
Photogrammetry matches surface appearance across views. A shiny surface shows reflections that move as the viewpoint changes, so the same physical point looks different in every photo and cannot be matched. A transparent surface shows whatever is behind it, so there is no stable surface detail at all. A light coat of matte spray or talc rescues mildly glossy test objects; it cannot fix transparent or mirror surfaces.
Can photogrammetry measure real dimensions?
Yes, but only after scaling against a known reference. Without one, the shape is correct and the size is arbitrary. With a calibration cube of known size scanned in the same session, the mesh can be scaled to millimetres and the error measured directly — and that measured error, not a catalogue figure, is the honest accuracy statement for a given build.
Is a phone camera good enough?
For learning and for display models, yes, provided the capture rules are followed: fixed tripod, locked focus, even lighting, and strong overlap. Phone sensors, lens distortion and fixed focus do cap fine detail compared with a dedicated camera, and that cap should be stated plainly in any report rather than hidden behind a claimed accuracy figure.
Limitations of this guide
This guide describes the standard photogrammetry pipeline at the level a student needs to build and debug a scanner. It does not cover advanced topics such as coded targets, multi-camera rigs, aerial survey calibration, or metric accuracy certification. Figures quoted for commercial scanners are reported launch figures, not measurements, and no accuracy figure in this guide should be quoted as a property of a student build. Every build must report its own calibration-cube error, measured, not assumed.
The turntable scanner build that applies this guide is here: https://projectech.in/projects/m1-ultra-inspired-turntable-3d-scanner-photogrammetry-esp32/