What Is Reprojection Error in Photogrammetry?
Reprojection error is the first number most photogrammetry reports show and the one most often misread. It measures how well the solved cameras and 3D points agree with what was actually seen in the photos. It says a lot about whether the alignment is healthy and nothing about whether the map is in the right place on the ground.
How reprojection error is measured
During alignment, the software finds the same feature in several photos and solves where each camera was, how its lens distorts, and where each of those features lies in 3D. Take a solved 3D point, project it back into one of the photos with that photo's solved position and lens, and it lands close to, but not exactly on, the spot where the feature was detected. The distance between the two, in pixels, is the reprojection error of that observation.
Bundle adjustment is the step that moves cameras, lens parameters and points to make the sum of these squared distances as small as possible. The figure in a report is usually the root mean square over all observations, sometimes with a breakdown per camera.
What a healthy value looks like
For a well-calibrated camera and sharp photos, a reprojection error below about one pixel is the usual sign of a sound alignment, and good aerial blocks are often well under that. A value several pixels high means something is wrong: bad matches, blur, a poor lens model or a camera that does not fit the rest of the block.
Always read the number together with the image scale it was measured at. If the engine aligned downscaled photos, an error measured at that scale looks smaller in pixels than the same error at full resolution. Stream Capture reports the reprojection RMS at full image resolution for that reason, beside the tie-point count, the ground sampling distance and the overlap.
What makes it rise
Wrong matches on repetitive texture, such as roof tiles, rows of crops or paving. Things that move between photos: water, vegetation in wind, vehicles, shadows on a long flight. Motion blur and noise. A lens model that does not fit, or a wrong focal length to start from. And on many consumer cameras, the rolling shutter, which gives each row of the image a slightly different camera position.
Why a low reprojection error is not map accuracy
Reprojection error is internal: it compares the model with the photos it was built from. A block can agree with itself to a fraction of a pixel and still sit metres away from its true position, be slightly wrong in scale, or be bent into a shallow bowl. That systematic curvature, known as doming, can appear with near-parallel downward photos and a self-calibrated lens, and it barely shows in the reprojection error.
Only points measured independently on the ground can tell you how accurate the map is. That is why a quality report should separate three numbers and never present one as another.
| Measure | Unit | What it tells you | What it does not tell you |
|---|---|---|---|
| Reprojection RMS | Pixels | The photos and the solved model agree with each other | Position, scale or height on the ground |
| GCP RMS | Metres | How well the model fits the control points it was adjusted to | Accuracy elsewhere; it is optimistic by construction |
| Check point RMS | Metres | The error at surveyed points the solution never used | Error in areas far from any check point |
How to use it
Use reprojection error as a health check and a comparison. Compare it between runs of the same data to see whether a change in settings helped. Look for single cameras with much higher error than the rest. Remove tie points with high error and let the cameras be recomputed if the software allows it. Then judge accuracy where it is decided: at independent check points.
Frequently asked questions
What is reprojection error?
The distance, in pixels, between where a solved 3D point projects into a photo and where the matching feature was actually detected in that photo. Reports usually give its root mean square over all observations.
What is a good reprojection error?
With a well-calibrated camera and sharp photos, below about one pixel is the usual sign of a sound alignment. Several pixels means something is wrong. Read it at the image scale it was measured at.
Why is my reprojection error high?
Common causes are wrong matches on repetitive texture, moving water or vegetation, motion blur, a poor lens model or wrong starting focal length, and rolling shutter on a moving camera.
Does a low reprojection error mean my map is accurate?
No. It shows the model agrees with its own photos. The block can still be shifted, scaled or slightly bowled; only independent check points measure accuracy on the ground.
Is reprojection error measured in pixels or metres?
In pixels, on the image. GCP and check point errors are in metres on the ground, which is why they answer a different question.
See reprojection, GCP and check point errors side by side in one report
Native Windows · fully offline · free beta