Troubleshooting Heritage Photogrammetry: Doming, Holes, Reflective and Repetitive Surfaces

Diagnose Before You Reprocess

The instinct when a model looks wrong is to reprocess at higher quality settings. This almost never helps, because the common failures are not caused by processing quality — they are caused by the input images or the capture geometry. Reprocessing bad input at high quality produces a high-quality bad result and wastes hours of processing time.

Diagnose first. Look at the alignment report: how many cameras aligned, what is the reprojection error, where are the failed cameras? Look at the sparse cloud before the dense cloud: doming, gaps and misalignment are usually visible in the sparse cloud in seconds, before you spend an hour on a dense reconstruction. Identifying which of the failure modes below you are facing tells you whether the fix is a processing adjustment, a re-capture, or an acceptance that the surface needs a different technique entirely.

Doming and the Bowl Effect

Doming (also called the bowl effect) is a systematic distortion in which a surface that should be flat is reconstructed as gently curved — bowed outward or dished inward, with the error greatest at the centre and least at the edges. On a facade it makes a flat wall bulge; on a drone survey of flat ground it lifts or sinks the centre of the site. It is the most important failure mode to understand because it is invisible to the eye in the viewport yet fatal to measured accuracy.

The cause is a subtle interaction between the camera's lens distortion model and a capture geometry with too little variation in viewing angle — classically, nadir-only drone imagery (all photos looking straight down) or square-on-only terrestrial imagery (all photos perpendicular to the wall). With no convergent (angled) images, the software cannot fully separate genuine lens distortion from surface shape, and the residual error accumulates as a systematic curve.

• Capture fix (the real fix): add convergent oblique images — photos taken at 20–45° across the surface, in addition to the square-on or nadir pass. For drones, add oblique passes and cross-strips; for facades, add angled images along the wall. This single practice eliminates most doming.
• Control fix: well-distributed ground control points, especially across the middle of the surface, constrain the geometry and remove doming even from a marginal image set — GCPs at the perimeter only are less effective against doming than GCPs distributed across the surface.
• Processing mitigation: after alignment, optimise cameras with the appropriate lens parameters enabled; a pre-calibrated lens (fixed, known distortion) reduces the software's freedom to introduce doming. This helps but does not substitute for convergent capture.
• Detection: doming is best caught with independent check points — measured points not used in the alignment. If check-point residuals grow toward the centre of the model, you have doming.

Holes, Gaps and Missing Recesses

Holes are areas where the model has no reconstructed surface — gaps in a facade, missing depths in carved recesses, blank patches on a sculpture. In heritage work, holes most often appear exactly where the detail is richest: deep in the recesses of carved ornament, in undercut sculpture, and in shadowed junctions, because these areas are the hardest to photograph from enough angles.

The causes are occlusion (the surface simply was not seen from enough viewpoints — the camera could not see into the recess), insufficient overlap (too few images covering the area), and featureless or shadowed surfaces (the area was seen but had no texture or light for matching). Each has a different fix, and identifying which applies is the key to solving it.

Reflective, Wet and Polished Surfaces

Structure-from-Motion assumes that a physical point looks the same from different viewpoints — this is what lets it match the point across images. Reflective, polished, wet and transparent surfaces break that assumption: a specular highlight or reflection moves as the camera moves, so the 'feature' the software matches is not a fixed point on the surface at all. The result is noise, holes, or spurious floating geometry on polished granite, glazed tile, wet stone after rain, glass, and metal.

This is a fundamental limitation, not a settings problem — no processing option makes a mirror photogrammetrically measurable. The fixes work by changing the surface's appearance at capture or by choosing a different technique.

• Diffuse the light: reflections and highlights are worst under hard directional light; overcast conditions, shade, or diffused artificial light dramatically reduce specularity on polished stone.
• Cross-polarisation: a polarising filter on the lens combined with a polarised light source removes specular reflections optically — effective for controlled close-range capture of polished or varnished surfaces.
• Wait for the surface to dry: wet stone after monsoon rain is far harder than dry stone; where possible, capture in dry conditions.
• Matting agents (with caution): removable matting sprays exist but must never be used on heritage fabric without conservation approval — most heritage surfaces prohibit any applied substance.
• Switch technique: for genuinely reflective or transparent elements — polished metal, glass — terrestrial laser scanning (with its own limitations on glass) or direct measurement is more appropriate than photogrammetry.

Repetitive and Featureless Patterns

Repetitive patterns cause a distinctive failure: the software matches a feature to the wrong instance of an identical-looking feature elsewhere on the surface. A wall of identical bricks, a repeating carved frieze, a tiled floor, or a row of identical balusters can all confuse feature matching, producing misaligned sections, ghosting (the same element appearing twice), or a model that folds onto itself.

Featureless surfaces cause the opposite problem — no distinctive features to match at all — and produce holes or failed alignment, as covered above. Repetitive surfaces have too many identical features; featureless surfaces have too few distinctive ones. Both are matching problems, and both are addressed by giving the software unambiguous geometry to lock onto.

• Maintain high overlap and small steps: with enough overlap, adjacent images share so much context that the software can disambiguate which instance of a repeating element it is looking at.
• Capture the wider context: include images that show the repetitive area within its surroundings (a corner, an opening, an edge) so the pattern is anchored to unique features nearby.
• Add distinctive control targets: placing a few coded targets or distinctive markers across a repetitive surface gives the software unique points to anchor the matching — highly effective for repeating friezes and tiled areas.
• Avoid photographing a repetitive surface in isolation: a frame containing only identical repeating elements and no unique reference is the worst case; always include something that breaks the repetition.

Alignment Failure and Failed Cameras

Sometimes alignment fails outright: a large fraction of cameras do not align, the model splits into disconnected pieces, or the sparse cloud is a chaotic mess. The alignment report is the diagnostic — cameras that failed to align show greyed out, and their pattern reveals the cause.

The usual causes are motion blur (blurred images have no sharp features to match), insufficient overlap (a gap in coverage disconnects the sequence), featureless images (frames of plain surface with nothing to match), and drastic changes in scale or lighting between frames (mixing very different distances or exposures in one chunk). Each is diagnosable from which cameras failed and what those images contain.

1. 1Open the alignment report and identify the failed (unaligned) cameras.
2. 2Inspect the failed images: are they blurred, featureless, badly exposed, or isolated with no overlap? The content reveals the cause.
3. 3Disable genuinely bad images (blurred, featureless) and re-run alignment — a few bad frames can prevent good ones from aligning.
4. 4For a coverage gap, there is no software fix — the missing overlap must be filled by re-capture; note it for the next visit.
5. 5For mixed scales/lighting (e.g. drone plus close-range in one chunk), split into separate chunks, align each, then merge — very different focal lengths and distances align poorly together.
6. 6Increase key point and tie point limits, or use higher alignment accuracy, only after ruling out the input problems above — settings rarely rescue genuinely poor input.

Surface Noise and Scale Errors

Two remaining problems are worth naming. Surface noise — a fuzzy, thick, or grainy reconstruction instead of a crisp surface — usually comes from soft images (slight blur or high ISO noise), weak overlap, or low-texture surfaces at the edge of matchability. The fix is at capture (sharper images, lower ISO, better light); in processing, gentle noise filtering in CloudCompare helps but cannot manufacture detail that the images did not resolve.

Scale errors — a model whose dimensions are wrong — arise when scale was set from unreliable references. A model scaled only from camera EXIF or a single measurement can be several percent out. The fix is to introduce and check scale properly: two or more scale bars or, for measured work, ground control measured by survey instruments, with at least one independent check measurement confirming the scale is correct across the whole model.

Quick Diagnostic Reference

Common Mistakes

• Reprocessing at higher quality to fix a bad model — quality settings do not fix capture-geometry failures like doming, occlusion holes or repetitive-pattern confusion.
• Not checking for doming because the model looks fine — doming is invisible to the eye; only check points or measurement reveal it.
• Presenting interpolated (hole-filled) geometry as measured record — filled holes are fabricated surface and must be disclosed, never passed off as survey data.
• Trying to photogrammetrically capture mirrors, glass or polished metal — these break the core assumption of SfM; use a different technique.
• Mixing drone and close-range images in a single chunk — very different scales align poorly; process in separate chunks and merge.
• Trusting scale from a single measurement or EXIF — always use redundant scale references and an independent check.
• Skipping the alignment report — it names the failed cameras and, through them, the cause of most alignment failures.

Professional Practice

In professional practice, the value of understanding these failure modes is that it converts a mysterious 'the model came out bad' into a specific diagnosis with a specific action — re-capture this recess from these angles, add convergent images to remove this doming, place targets on this repeating frieze. It also feeds directly back into capture discipline: a practitioner who has diagnosed doming once captures convergent images automatically forever after.

The most important professional habit is verification before delivery. Because the most dangerous failure — doming and scale error — are invisible in the model and only appear under measurement, a professional heritage deliverable includes independent check-point residuals and a confirmed scale check. A model delivered without these has an unknown accuracy, and 'looks correct' is not an accuracy statement.

Finally, knowing the failure modes shapes honest scoping. Some heritage surfaces — highly reflective polished granite, dark undercut sculpture, glass — are genuinely poor photogrammetry subjects. Recognising this at the quoting stage, and specifying laser scanning or a hybrid approach where appropriate, prevents the far worse outcome of promising a photogrammetric result on a surface that cannot deliver one.