Photogrammetry recovers three-dimensional geometry from a set of overlapping photographs. Given enough images of the same object from enough positions, the shape of that object and the position of every camera can both be solved for, without any of that information being known in advance.
It is one of the genuinely impressive results in applied vision, it has been in industrial use for decades, and it is routinely chosen for jobs that did not need it. The output is compelling: a textured mesh that can be rotated, sliced and measured. That compulsion is precisely the risk, because a reconstruction that took three days of capture and a night of compute is frequently used to extract a single dimension that a tape measure or an existing dataset would have supplied in minutes.
Understanding when the reconstruction is the point, and when it is an elaborate route to a number, requires knowing what the pipeline is actually doing and what it costs.
What the pipeline is doing
The process has a consistent shape regardless of the software.
Feature detection finds points in each image that are distinctive enough to be recognised again from a different viewpoint: corners, texture patches, anything with local structure. Smooth regions produce nothing, which is the origin of most later trouble.
Feature matching finds which features in one image correspond to which in another. This is the fragile step. Correspondence is established on appearance, and anything that makes two different places look alike produces wrong matches.
Pose estimation, usually called structure from motion, solves simultaneously for where each camera was, how it was oriented, and where the matched points sit in space. It works incrementally or globally, and it produces a sparse cloud of points plus a camera for each image.
Bundle adjustment then refines everything at once, minimising the discrepancy between where points project into the images and where they were observed. This is what turns a rough solution into a consistent one, and it is where accumulated drift gets distributed rather than eliminated.
Densification, or multi-view stereo, uses the now-known camera geometry to compute depth for far more pixels than the sparse features covered, producing a dense point cloud.
Meshing converts points into a surface, and texturing projects image colour back onto that surface. These last steps are largely cosmetic in engineering terms and are what makes the result look like a model rather than a measurement.
A crucial point about the whole chain: without external information, the result is correct only up to an unknown scale. The reconstruction knows the scene's shape and the cameras' relative positions, not how big any of it is. Scale comes from ground control points, from a survey-grade positioning system on the camera, or from an object of known dimension placed in the scene. A reconstruction without one of those is not measurable in real units, however precise the mesh appears.
Overlap and baseline
Capture geometry determines whether reconstruction succeeds, and it is decided before any software runs.
Overlap is how much of the scene appears in consecutive images. Aerial and drone work conventionally uses high forward overlap along the flight line and substantial side overlap between lines, and the reason is not redundancy for its own sake: every surface point needs to be visible in several images from meaningfully different angles for its depth to be constrained. Points visible in only two images are weakly determined and points visible in one are not reconstructed at all.
Baseline is the distance between camera positions. It creates the parallax that carries depth information, and it sits in direct tension with matching. A wide baseline gives strong depth sensitivity and makes surfaces look different between views, which makes matching harder. A narrow baseline matches easily and determines depth poorly. The workable range is bounded on both sides, and the ratio of baseline to distance from the subject is the quantity that matters rather than the baseline alone.
Two capture patterns fail predictably. A straight line of images along a facade with no lateral variation gives a degenerate geometry with a poorly constrained solution. And images taken all from roughly the same point, rotating rather than translating, provide no baseline at all: that arrangement produces a panorama, not a reconstruction, and no processing will extract depth from it.
Vertical structure needs vertical coverage. Reconstructing a building from nadir imagery alone yields good roofs and poor walls, because the walls are barely visible. Oblique imagery from several directions is what makes facades work, which is why building reconstruction flights use oblique passes rather than a simple lawnmower grid.
What makes a scene hard
Failure in photogrammetry is not gradual. Scenes tend to reconstruct well or to collapse, and the causes are identifiable in advance.
Texture-less surfaces give no features. Smooth painted walls, new concrete, calm water, clear sky, fresh snow and uniform metal roofing produce nothing to match, and the reconstruction has holes exactly where those surfaces were. Software fills them by interpolation, which looks like geometry and is invention.
Repetitive patterns produce confident wrong matches. Regular window grids, tiled facades, brickwork, fencing, orchard rows and identical panels all look the same at multiple offsets. Matching latches onto the wrong repeat, and the resulting pose error can distort a whole section of the model or split one wall into two parallel surfaces.
Reflective and transparent surfaces move as the camera moves. A window's reflected content changes with viewpoint, so features detected on it are not attached to any fixed point in space. Glass facades, polished stone, wet surfaces and vehicles all contribute features that actively mislead the solution rather than merely failing to help.
Moving objects violate the core assumption that the scene is static. Pedestrians, traffic, foliage in wind and shadows drifting across the day all appear in different places in different images. Robust estimation discards some of it, and dense reconstruction turns the rest into fragments and ghosting.
Changing light is the underrated one. Capture spanning several hours means shadows have moved, so the same surface has different appearance and, worse, shadow edges are strong features that get matched as if they were physical. Overcast conditions are preferred for exactly this reason: soft shadows, consistent illumination, no blown highlights.
Thin structures are lost. Railings, cables, poles and branches are too small to be reliably matched from multiple views and disappear or become noise. For clearance and obstruction work, that is the survey's whole purpose evaporating.
Scale extremes strain the process from both ends. Very large sites accumulate drift that bundle adjustment can only partially resolve without control points. Very small or very close subjects run into depth of field limits, where much of each image is out of focus and features are unreliable.
The cost is capture effort first, compute second
The compute is easy to see and easy to budget. Dense reconstruction is expensive but it is a machine cost that scales with images, and it can be scheduled.
The cost that dominates is capture. Photogrammetry requires physically being at the site with a plan: coverage designed in advance, consistent conditions, discipline about overlap, and control points measured if the result must be in real units. That takes people and time, it is weather dependent, and it is often permission dependent.
The expensive part of that cost is the return trip. Whether the capture is adequate is frequently not knowable until processing completes, by which time the crew has left. A gap in coverage, a poorly lit facade, a degenerate camera path, an unnoticed change halfway through: each means going back. The cost of a photogrammetry project is better modelled as the capture cost multiplied by the expected number of visits than as a single visit plus compute.
There is a storage and handling cost that is easy to underestimate. Dense clouds and textured meshes are large, awkward to version, awkward to serve, and awkward to join against ordinary geospatial data. A reconstruction that lives in one workstation's project folder is not an asset.
What accuracy a reconstruction actually has
A finished mesh invites measurement to whatever precision the software will display, and the displayed precision has no relationship to the accuracy of the underlying geometry. Three distinct quantities get conflated, and separating them is necessary before any figure derived from a model can be defended.
Resolution is how finely the surface is sampled, driven by ground sample distance in the imagery. It sets the smallest feature that can appear at all. A model built from imagery where one pixel covers several centimetres will not contain anything smaller than that, regardless of how dense the point cloud is, because densification interpolates between observations rather than adding information.
Relative accuracy is how correct distances are within the model. This is usually the best of the three and is what most users implicitly assume they are getting. It depends on capture geometry and on how well bundle adjustment converged, and it degrades with distance across the model as drift accumulates.
Absolute accuracy is how correct the model's position and scale are in the real world. This depends entirely on the control: on how many known points were measured, how accurately, and how they were distributed. Control clustered in one corner of a site leaves the far corner unconstrained, and the model can be internally consistent while being wrong in scale and orientation across the extent.
The failure this produces is subtle. Someone measures a span across a large reconstruction and quotes it to the millimetre. The model is internally excellent, so a short local measurement would be sound, but the long span has accumulated drift and inherited a scale set by control at one end. The number is precise and wrong, and nothing in the interface indicated which kind of measurement was being taken.
The corrective habit is to hold back at least one measured control point from the solution and check it afterwards. A point that was not used to constrain the model, and that lands where it should, is real evidence about absolute accuracy. Every point used in the fit is not.
Alternatives that are frequently sufficient
Before committing to reconstruction, it is worth asking what the output is really for, because several cheaper routes deliver the same answer.
Existing footprints plus a height estimate give building volume and rough massing without capture at all. For anything at city scale, this is the standard approach, and a reconstruction would not improve the decisions that follow it.
Single-image depth estimation gives approximate geometry from photographs already held, with an honest error band. Where the question is relative arrangement rather than precise dimension, it answers it at zero capture cost.
Rectification of a single facade photograph gives accurate two-dimensional measurements on a plane, which is what most facade questions actually require. If the deliverable is a window schedule or a facade area, a rectified image with a known reference dimension is faster, simpler and more accurate than measuring off a mesh.
Where laser scanning data already exists, whether airborne survey or a terrestrial scan, it provides geometry directly and with known accuracy. Photogrammetry competes with it on cost of acquisition, not on quality of result.
And direct measurement remains the cheapest option for a small number of dimensions on an accessible site. A visit with a laser distance meter answers a handful of specific questions definitively in an afternoon.
The decision rule
Reconstruction is genuinely needed when the deliverable is the geometry itself, when many measurements will be taken from it over time, or when the surface is irregular enough that no simpler representation captures it.
The specific cases that justify it:
- The output is a model that others will interact with, measure from, or use as a base for design work. Here the mesh is the product, not an intermediate.
- The measurements required are numerous, not yet fully specified, or will be revisited. Capturing once and querying repeatedly beats returning to site for each new question.
- The surface is complex and irregular in a way that a footprint plus a height cannot represent: heritage structures, damaged buildings, stockpiles, terrain, anything organic.
- Volume is the quantity of interest. Volumetric measurement of an irregular mass is the case where reconstruction has no cheap substitute.
- The state must be recorded for later comparison. Condition surveys, deformation monitoring and construction progress all rely on comparing reconstructions across time, which requires the reconstruction to exist.
And the counter-rule, which is the one worth holding onto: if the deliverable is a small number of scalar values, and any of those values is obtainable from an existing dataset or a direct measurement, photogrammetry is the expensive path to a number that was already available.
If the reconstruction is going ahead
The projects that fail do so at capture, and the failure is almost always discovered in processing. Some discipline at capture time removes most of it: plan coverage explicitly rather than trusting judgement in the field, capture more overlap than seems necessary since it costs storage and not much else, add oblique views for anything vertical, work in consistent light and finish within a window short enough that shadows have not marched across the subject, place and measure control or at minimum a scale bar so the result is in real units, and process a rough preview before leaving the site.
That last item is the highest value habit in the whole discipline. A fast, low quality reconstruction run on the spot will not be the deliverable, but it will reveal the hole in coverage while the crew is still standing in front of the building. Every other check can happen later. That one cannot.