Photogrammetry on Site: Where Model Accuracy Is Lost
Plausible geometry and accurate geometry are not the same thing.
Reconstruction from images almost always looks convincing, but the discrepancy with reality comes to light during measurements: the facade plane “breathes,” dimensions drift by centimeters, curved sections produce waviness.
Most of the error is introduced at the site, not in photogrammetric processing from photos: frames go into Agisoft Metashape, but you cannot extract from them what is not there.
The first cause is a weak survey plan: a single pass around the perimeter, frame overlap below 60%, and constant altitude. There is not enough parallax, the frame tie relies on isolated points, and on flat surfaces gaps and noise appear.
The second is the absence of survey control points: the model matches in shape but drifts in scale and orientation, and this is visible only when checked against control targets.
The third group of object survey errors is the physics of the frame. Glare on glass and polished surfaces produces false matches, moving people and machinery produce “double” geometry, and monochrome areas without texture tie poorly. Shadows captured from different angles are read by the algorithm as relief.
These losses are cheaper to prevent than to compensate for: holes in the cloud take longer to fill, and part of the site has to be re-surveyed. That is why 3D model accuracy is planned before going to the site — passes, overlap control, control and check points.
Measurable Effect of 3D Reconstruction from Photographs
Switching to photogrammetry changes not the tool but the very order of working with the site. Instead of a crew of surveyors with tape measures and laser rangefinders — a walk-through or flyover of the site and surveying along pre-calculated routes with frame overlap.
Measurement timelines shrink from several days to several hours on site plus overnight processing, and the result immediately goes into a 3D model of the site, not into scattered sketches and tables.
Next, the economics of error changes. Previously, each project discipline — architecture, structures, utilities — built its own geometry from its own measurements, and discrepancies surfaced on site as cut masonry and relocated risers.
Now everyone works from a single site model: clashes are caught before documentation is issued, not after. This is what provides the main reduction in rework — the share of work redone due to inconsistencies in source data falls by a factor.
Separately, about the quality of source data.
Surveying with 70–80% overlap and control measurements along baselines give convergence within a centimeter, while processing and alignment in Agisoft Metashape produce a reproducible result: repeat surveying yields the same geometry, which allows model versions to be compared and changes to be seen. The completeness of source data no longer depends on who managed to measure what.
The number of site visits is reduced from three or four to one: repeat visits are needed only for concealed works or to clarify areas hidden by equipment.
Measurement timelines cease to be a bottleneck at the start of the project, and updating the model after changes becomes a routine operation rather than an emergency visit.
Ground and Aerial Photogrammetry: What to Choose
The choice of approach is determined not by the size of the site but by which of its surfaces are not visible from the selected points.
Ground photogrammetry is a walk-around of the site along circular routes with frame overlap: it captures facades and vertical planes but does not cover the roof, upper tiers, and inner courtyards.
Aerial surveying covers the plan and roof, but under tree canopies, awnings, and in narrow passages between walls, frame connectivity breaks, and holes remain in the model.
Interior surveying is done differently: movement inside premises, wide angle, constant light — otherwise monochrome walls and partitions do not find corresponding points, and the reconstruction “drifts.” The hybrid approach stitches both sets using shared control points, while laser scanning adds coordinate support where texture-based photogrammetry cannot cope.
| Option | When It Fits | Limitations |
|---|---|---|
| Ground photogrammetry | free-standing sites with open visibility, facades, quarries, monuments | roof and upper planes inaccessible; backlight and deep shadows cause gaps |
| Aerial surveying | buildings, roofs, industrial sites, linear objects | dense vegetation, canopies, narrow passages between walls |
| Interior surveying | premises, workshops, utility assemblies | monochrome and glossy surfaces, dead ends, small frame area |
| Hybrid approach | complex-shaped site where both interior and exterior are needed in one model | requires a unified control point scheme; more field work |
| Laser scanning | high-precision measurements, weakly textured surfaces, support for photography | less color information; needs tying to photo scale |
In practice, the decision is made using a map of non-visible zones: where there are more of them, a flyover, scanning, or interior pass is added. This kind of analysis at the planning stage eliminates the need to re-survey the site.
How the Project Runs: From Survey to Verified Model
The project starts with a task audit: it is important to understand what the model is needed for — measurements, design, visual control, or catalog population.
This determines the required accuracy, detail, and which objects are surveyed separately and which can simply be walked around along the perimeter.
- Task audit. We record object dimensions, permissible error, and result format: point cloud, polygon mesh, or orthophoto.
- Survey plan. We calculate frame overlap, number of tiers and angles, camera stations. We separately plan control points and scale markers if the model is tied to real coordinates.
- Image processing setup. We reject blurred frames by sharpness and metadata, set camera calibration and surface reconstruction density.
- Model accuracy control. We compare the result against independent control measurements: discrepancies show where the geometry “drifts” and supplementary survey is needed.
- Delivery to the client. We hand over the model in agreed formats together with an accuracy report and source images.
- Team training. We show how to conduct surveys on new sites so that the result remains reproducible without us.
If at the control stage discrepancies exceed tolerance, the cycle is partially repeated — supplementary surveying of problem zones costs less than a full re-survey of the site.
What You Receive: Model, Plans, Regulations
The deliverable is not a single model file but a set that lets the site live on without the survey author. Everything needed for acceptance, repeat surveying after a season, and transferring data into another design system is described and sits alongside the geometry. Below is what exactly you receive.
- A ready model broken down by layers: terrain, buildings, vegetation, and utility objects are separated so they can be turned on and off in a GIS or design system. Alongside are a textured polygon mesh, point cloud, orthophoto, and elevation map.
- A build quality report: georeferencing residual, point density, share of overlapping coverage, and a gap map. It shows where the geometry is reliable and where supplementary survey is needed.
- Survey and georeferencing plans: flight paths, station coordinates and elevations, positions of ground control markers, and the coordinate and elevation system used.
- A survey regulation: requirements for frame overlap, lighting, scale, and weather limits. Another operator can repeat the survey from it and get a comparable result.
- Source images with metadata: frames with coordinates, angles, and camera parameters remain with you, so the model can be rebuilt with different settings.
- Model documentation: layer descriptions, file structure, achievable accuracy for each object type, and instructions for loading into your system.
- Model access rights: role separation for viewing, editing, and export, the procedure for transferring data to third parties, and terms of use.
We deliver the set with a closed access loop: you decide which contractors and departments see the geometry and who sees only the final plans. The plans and regulations are deliberately written so that an implementer can read them without our support.
Case Study: Reconstruction of an Industrial Facility from Photos
Before: Manual Measurements in Dense Development
At an operating production facility, measurements were taken manually: a crew of three walked around workshops and pipe racks with a tape measure and total station.
On six thousand square meters, this took three weeks, and actual accuracy stayed around ±40 mm — at interfaces with existing structures and in assemblies with dense routing, this was not enough, and some points had to be re-measured on a repeat visit.
The basis of the new model was assembled with photogrammetry: about 4,800 images with overlap across the entire area; coordinate referencing was provided by 24 control points measured with a total station.
Processing was carried out in Agisoft Metashape, yielding a point cloud, orthophoto, and a complete three-dimensional model in a geodetic coordinate system.
The survey was coordinated with workshop operations: the walk-around was done during process windows, and equipment was not stopped.
Metrics Before and After
Model accuracy against control points was 8–12 mm versus ±40 mm for manual measurements; spread on check spans did not exceed 15 mm. The field stage was reduced from three weeks to four days: surveying took one shift, the rest was processing and control.
Repeat visits for measurements were no longer required, and clashes that previously surfaced during installation are now visible on the model: the designer gets a section and elevation for any assembly.
What to Do If the Site Has Already Been Partially or Poorly Surveyed?
We start with an audit of existing materials: we check image metadata, actual frame overlap, edge sharpness, lighting uniformity, and the presence of control points.
Already at this stage it is clear whether there is enough data for a dense cloud and elevation map, and where an angle was missed — in most projects the deficit is covered by supplementary surveying, without a repeat walk-around of the site.
Supplementary surveying works when the coordinate system is preserved: control markers on the site, image geotags, or referencing to an already assembled model.
Then the new block of images is placed into the same space, and both parts are merged into one scene in Agisoft Metashape — surfaces are joined by shared areas, not by our assumptions.
We separately monitor consistency of light and optics: changing the lens or time of surveying creates a seam that is later visible in the texture.
Re-surveying is needed where fixing is more expensive than shooting anew: overlap below the working threshold, frames blurred by motion, part of the site occluded, a surface that is monochrome and without relief, metadata lost, or surveying done with different cameras at different distances.
Stitching such sets produces gaps and “blur” that cannot be fixed by settings — time is spent, and the result remains incomplete.
We make the decision using a coverage map: we mark zones where there is enough data, where supplementary surveying helps, and where the model cannot be assembled without a new walk-around.
The client receives a list of specific areas and the scope of work for each — this way a partially surveyed site does not turn into a project from scratch.
Technical Concerns: Accuracy, Timelines, and Model Integrity
Concerns about photogrammetry usually come down to four questions: what accuracy is achievable on a specific site, how survey conditions affect it, whether archival images are suitable, and whether the model will open in your CAD system. We answer briefly — before work begins.
What accuracy of 3D reconstruction can realistically be achieved?
Accuracy is set not by the algorithm but by the source material: frame overlap, sharpness, and the geometry of the walk-around.
In a survey with about 70% overlap and even light, error stays within 1–2 mm per meter of dimension — this is enough for measurements and geometry control. For fine sculpture and relief surfaces, we plan additional passes.
How critical are lighting and weather?
Light affects point cloud density and texture, not geometry. Hard shadows and glare on glass and metal leave “holes” — we close them with additional passes or surveying in overcast conditions. Areas that could not be surveyed cleanly are marked in the quality report.
Are archival images suitable?
Yes, if there is overlap between frames and the camera type is known. We also accept film scans and phone images: we calibrate from the frames and restore their relative position.
Accuracy will be lower than controlled surveying, but it is sufficient for documentation and visualization of a lost object.
Will the model open in your system?
We deliver a point cloud, polygon mesh, and, if needed, parametric geometry for CAD. Before handover, we check import and send a sample — you test it on your side. We agree on a non-standard format before starting.
Let’s Discuss the Task: What to Prepare for Assessment and a Work Plan
To come back with a stage plan and project timelines rather than a wide range, we need a minimum of source data. Below is what most strongly affects the assessment of a photogrammetry task; we will clarify the rest in conversation.
- Survey set: a series of site images with frame overlap and no gaps in complex areas — it shows whether a connected model can be assembled.
- Metadata: preserved EXIF, georeferencing, control points or geodetic markers — they determine whether a point cloud in real coordinates can be produced.
- Reference result: an example model or orthophoto that you consider suitable — it fixes requirements for detail and texture.
- What is needed at output: point cloud, dense mesh, textured model, orthophoto, sections — this determines the composition and number of stages.
- Coordinate system, units of measurement, and formats for delivering the result (OBJ, PLY, LAS, GeoTIFF) — so the model fits directly into your tool.
- Constraints: restricted zones, access, survey regime, data volume — affect the order of work and timelines.
- Who on the client side accepts the result and by what criteria — this is the acceptance point at each stage.
Based on these points, we assemble a work plan: stages, the set of checks at each, preliminary project timelines, and places where your decisions are needed.
Describe the task and attach what you already have: even a partial data set provides a timeline reference; we will clarify the rest in conversation.
