Judging Capture Quality

The first question about any capture is the simplest and the most often skipped: is the thing you need to measure actually present in the data? A scanner or camera only records what it can see from where it stood, so every dataset is shaped by line of sight. Behind every column, above every suspended ceiling, inside every reveal and under every pipe run there is a shadow — a region the beam or lens never reached — and in that shadow there are simply no points. The cloud does not announce these gaps; it just quietly has nothing there, and a surface modelled across a gap is a guess wearing the costume of a measurement.

Occlusion is the technical name for this, and it is inevitable, not a defect. The professional response is not to demand a gapless scan — there is no such thing — but to check that the gaps fall where they do not matter and that the surfaces you will actually design against are well covered from more than one direction. A wall seen from a single oblique angle is thin and unreliable at its edges; the same wall seen from two or three scan positions is solid and trustworthy. When you receive a dataset, the first pass is a coverage audit: open it, navigate to each element the brief depends on, and confirm there are real, dense points there — not a smeared skin bridging an absence.

Coverage failures are the quiet killers because they do not look like errors. A floor plan sliced from a cloud with a missing return wall still produces a clean line; it is just the wrong line, interpolated across nothing. Ceiling voids that were never reached, the tops of high parapets, the backs of machinery, rooms that were locked on the day — these are where the assumptions you were trying to eliminate creep straight back in. So when you judge a dataset, ask of every critical surface: was this *measured*, or is the software filling in? Good capture planning (Module 7) anticipates occlusion with extra stations and multiple vantage points; good quality judgement confirms, after the fact, that the plan worked where it counted. Coverage is the foundation check — if the data is not there, no amount of accuracy elsewhere can save it.

Zoom into any real point cloud and a wall is never a perfect geometric plane; it is a thin *slab* of points scattered a little to either side of the true surface. That scatter is noise — the random spread of individual measurements around reality, caused by the instrument, the range, the angle of incidence and, above all, the surface itself. A matte plastered wall returns crisp points; glass, polished stone, water, dark or glossy surfaces and very oblique hits scatter the beam and produce fuzzy, thickened, or outright ghost points floating in space. Learning to read noise is learning to see, at a glance, where a dataset is solid and where it is mush.

A trustworthy cloud reads as clean: surfaces are thin and well defined, edges are sharp, and when you slice a section the walls are clear lines rather than blurred bands. A flawed one shows thick fuzzy planes, speckle and stray points, doubled or smeared edges, and confident-looking geometry over exactly the shiny or distant regions where the method struggles most. You do not need a metric to start — the eye catches a great deal — but the honest move is to slice a thin section through a wall and literally look at how thick the band of points is, because that thickness is a direct picture of the noise you will be modelling against.

Density is the companion property: how many points per unit of surface, and therefore how fine a feature the cloud can actually resolve. A sparse cloud might be fine for gross room dimensions but unable to resolve a moulding profile, a conduit, or the edge of a rebate — the detail is simply not sampled. Density is not a virtue to maximise blindly, though: more points mean heavier files and slower work, and a cloud far denser than the job needs is its own kind of poor fit. The judgement is *fitness for purpose* — enough density to resolve every feature the deliverable must represent, at the noise level the tolerance can absorb, and no more. Read noise and density together and you can say, honestly, what this dataset is good for and what it is not.

A building is rarely captured in one go. It is recorded as many overlapping scans or photo sets, which software then stitches into a single cloud — a process called registration (Module 1 and Module 5). The danger is that registration introduces its own error, and this error behaves treacherously: a cloud can look razor-sharp at any one scan position while being subtly *distorted overall*, because small misalignments at each join accumulate down a corridor or along a facade. Local crispness is not global accuracy, and confusing the two is one of the most expensive mistakes in the field.

The visible symptom of bad registration is the doubled surface: one real wall appearing as two slightly offset planes, a blurred or split feature, a floor that steps where two scans disagree. Learn to hunt for these. Slice a section right across a span that was built from several scans and look for walls that have gone fuzzy or split; pan along a long element and watch whether it stays single and straight or drifts. But the eye is not enough, which is why a professional dataset comes with a registration report stating the cloud-to-cloud error — the residual disagreement between overlapping scans. A small, reported residual is evidence the stitching held; a missing report, or a large one, is a warning.

Registration tells you the scans agree *with each other*; it does not tell you they agree with *reality*. For that you need alignment to control — tying the capture to points whose positions are independently known, established by a surveyor or a measured control network (Module 1.4). Without control, a cloud can be internally consistent and globally wrong: the right shape at the wrong scale or orientation, or drifted over distance with nothing to pin it down. So the quality question has two layers. First: do the scans agree with each other (low registration residual)? Second: are they tied to something of known position, with the control residuals reported and within the stated accuracy? A dataset that can answer both, with numbers rather than adjectives, is one you can trust at the scale of a whole building — not just in a flattering close-up. And where that control must itself be survey-grade or legally binding, it belongs to a licensed surveyor, not to an assumption.

Every check so far can be gamed by a convincing-looking dataset, which is why the habit that underpins all of them is the oldest one in measurement: take an independent check. Before you commit a design to a captured dataset, measure something in the real world by a different means — a tape across a known length, a laser disto between two clear points, a door height, a diagonal — and compare it to the same dimension pulled from the cloud. If they agree within the accuracy you were promised, your confidence is earned. If they do not, you have caught a problem before it became a site failure, which is the entire point.

Choose check dimensions deliberately. A single short measurement proves little; a *long* one — the full length of a corridor, a building diagonal, floor-to-floor over several storeys — is far more revealing, because it is exactly over long distances that registration drift and scale error show up. Check in more than one direction, and check the things the design actually depends on. Record what you did and what you found: "corridor measured 12.184 m by disto, 12.19 m in cloud, within tolerance" is a sentence that protects you, your client and the project. This is not bureaucracy; it is the difference between a number you can defend and a number you merely hope is right.

The deeper discipline here is a mindset: verify, do not trust. A point cloud is millions of estimates, each carrying error, stitched and scaled by software — it is an instrument reading, not ground truth, no matter how photographic it looks. Treating it as self-evidently correct because it is detailed and beautiful is precisely the error this module exists to prevent. So build the reflex: for every dataset you will design against, confirm coverage of the surfaces that matter, read the noise and density, look for doubled surfaces and read the registration and control residuals, and take at least one independent check measurement over a long span. Where the stakes are high, where accuracy must be certified, or where the result is legally binding, that verification is a licensed surveyor's responsibility, not yours to improvise. Capture confidently; but sign off on nothing you have not checked.