Understanding the Point Cloud

Strip a point cloud down to its smallest unit and you find a single point: a record of one measured location in space. At minimum it carries a position - three numbers, x, y and z - describing where, relative to some origin, a laser pulse or a photogrammetric ray found a surface. That is the irreducible core. A point says, in effect, "there was something solid here, at this spot." Nothing more. It does not know what it hit, whether that something was a wall or a parked scooter, or whether the neighbouring points belong to the same object. It is just a measured coordinate.

Most points carry more than position. If the scanner has a camera, each point also gets a colour - red, green and blue values sampled from a photograph taken at the same time - which is why a good cloud can look startlingly like the real place. Laser scanners record an intensity value too: how strongly the pulse reflected back, which depends on the surface's material, colour and angle. Intensity is why a cloud viewed in greyscale still reveals lettering on a sign or the grain of a material even before colour is added. And after processing, points can carry a classification - a label saying this point is ground, that one is vegetation, this one is a building - though, crucially, that label is inferred later, not measured at capture.

Now gather millions - often hundreds of millions, sometimes billions - of these records together, each sitting at its true measured position, and you get the cloud. Seen from a distance the individual dots merge into continuous-looking surfaces, and the real geometry of the place emerges: walls, floors, columns, the sag of a beam, the profile of a cornice. But it is an illusion of continuity. Zoom in and the surfaces break back apart into the separate dots they always were, with empty space between them. This is the first and most important fact about a point cloud: it is discrete, not continuous. There is no surface between the points - only the points themselves, and your eye filling in the rest. Every later decision, from meshing to modelling, is in some sense about how to bridge the gaps between points responsibly.

If a point cloud is made of discrete points, then the obvious question is: how close together are they? That is density, usually described by point spacing - the typical distance between neighbouring points on a surface. A cloud with 5 mm spacing has a point every few millimetres; a cloud with 50 mm spacing leaves fist-sized gaps. Density is the resolution of your captured reality, and it governs what detail the cloud can possibly contain.

The consequences are concrete. At coarse spacing, a flat wall captures fine, but a delicate moulding, a thin conduit or a hairline crack simply falls between the points and is never recorded - it does not appear faint, it is absent. At fine spacing, those same features are resolved clearly. So the density you capture sets a hard ceiling on what you can ever measure or model from the data. This is why planning the scan matters so much (a later module): you choose a density for the smallest feature you care about, because you can always discard points later, but you can never invent detail you did not capture.

Two honest complications. First, density is not uniform. A scanner samples densely near itself and sparsely far away, so a single scan is dense by the instrument and thin across the room; only by combining many scan positions do you get reasonably even coverage. Photogrammetry and mobile scanning have their own uneven patterns. Second, more density is not automatically better. Every extra point is more data to store, move and process, and past a point it buys nothing but weight - a wall does not become more accurate for being sampled every millimetre instead of every five. The skill is matching density to purpose: survey of a carved heritage screen needs fine spacing; a warehouse shell needs far less. Throughout this course, treat any spacing figure as illustrative - real achievable density depends on the instrument, the range, the settings and the surface, and the authoritative numbers come from the equipment specification and, for anything binding, a surveyor. What matters conceptually is the principle: density is the resolution of your reality, set once at capture, and it quietly decides what questions the data can answer.

Here is where most misunderstandings live. A raw point cloud, straight from registration, is genuinely powerful - and genuinely limited, and the two facts sit side by side.

What you can do with raw points is a lot. Because every point is a true measured coordinate, you can measure any distance by picking two points - a room's width, a ceiling height, the clear opening of a door. You can slice the cloud with a plane to get a section or a plan cut at any level, revealing the true profile of walls and floors. You can compare the cloud against a design model to find deviation - where the as-built drifts from the as-designed, invaluable for construction verification. You can fly through it, take coordinates, and trace over it to draw. All of this works directly on the points, with no modelling at all. For many jobs - a quick dimensional check, a section, a deviation heat-map - the raw cloud is the deliverable.

What you cannot do natively is treat the cloud as objects. This is the crucial limit. The cloud contains no walls, no doors, no pipes - only points. No point knows it belongs to a wall; the cloud has no concept of a wall at all. So you cannot click to select "the wall" as one thing, you cannot ask for the area of a room or the volume of a space, you cannot schedule elements, and you cannot move or edit a component - because there are no components, only a dense fog of coordinates. The meaning - "these points are a wall, 230 mm thick, from floor to this ceiling" - is not in the data. It has to be added by a person (or, increasingly, by imperfect automation), and that act of adding meaning is exactly what scan-to-BIM is, the subject of Module 6.

Hold both halves together. The raw cloud is a measurable three-dimensional record of where surfaces are - superb for measuring, sectioning and checking. It is not an intelligent model, and expecting it to behave like one is the classic beginner's error. A point cloud answers the question "where is the surface?" brilliantly. It cannot, on its own, answer "what is this, and what are its properties?" - and knowing the difference is half the discipline.

A point cloud is measurement, so it carries all the imperfection of measurement - and a realistic mental model includes the mess, not just the clean ideal.

Every cloud contains noise: points scattered slightly off the true surface. A perfectly flat wall does not capture as a perfect plane but as a thin fuzzy slab, because each measurement has a small error and the points spread through it. Noise comes from the instrument's own precision, from the distance and angle to the surface, and from difficult materials - shiny, dark, wet or transparent surfaces misbehave, and glass and polished stone can produce points that are simply wrong. Alongside noise come outliers: stray points sitting well away from any real surface, caused by a beam catching a dust mote, splitting on an edge, or reflecting off a mirror so a 'ghost' surface appears where nothing exists. And the cloud records the moment of capture, so it is full of transient reality that is not the building at all: people walking through, parked vehicles, furniture, plants, a van at the kerb.

None of this is a fault to be embarrassed about; it is simply what captured data is. But it has consequences you must carry in your head. The thin fuzzy slab means a single picked point may be a millimetre or two off the true surface - fine for most architecture, but a reason to fit planes to many points rather than trust one click for anything precise. The outliers mean the raw data is not yet trustworthy for clean measurement until it is processed. The transient clutter means the cloud you open is not yet the building - it is the building plus everything that happened to be there. All of this is why the next lesson exists: registration and cleaning are the work of turning this honest mess into a usable dataset.

So hold this mental model, and let it replace any idea of the scan as a flawless digital twin: a point cloud is a dense collection of individual measured estimates of where surfaces are, each carrying a little error, unevenly spaced, peppered with noise and outliers, including things that are not the building, and containing no objects or meaning of its own. That sounds like a list of weaknesses. It is actually a list of the things a professional knows and accounts for - and it is exactly what separates someone who can be trusted with captured data from someone who will be fooled by it.