What a point cloud is and how it becomes a usable model

TL;DR: A point cloud is the raw output of a 3D scan: millions of measured coordinates on a part's surface, with no surfaces, edges or features attached. On its own it's a picture made of dots. This post covers what each point holds, why the cloud needs cleaning and aligning before anyone can use it, the three things it can turn into, and which file format to send us.

A point cloud looks like the finished job. Spin it on screen and the part is right there, every curve and corner, solid from a distance. Zoom in and it falls apart into dots with gaps between them.

That distance between how a point cloud looks and what it can do is where most confusion about 3D scanning lives. People send us a cloud and ask us to print it. Others get one back from a scanning bureau, open it, and can't work out why their CAD software won't let them change a dimension.

The cloud isn't the deliverable. It's the measurement, and measurements need interpreting before anyone can build from them. A single instrument like the Leica RTC360 records up to two million points a second. Nobody hand-checks two million of anything.

Here's what sits inside a point cloud, what has to happen to it, and what it can become.

What is a point cloud?

A point cloud is a set of measured points in three dimensions, each one recording a spot on the surface of a scanned object. Every point holds an X, Y and Z coordinate, and often colour and light intensity as well. The points aren't joined to each other, so the cloud has no surfaces and no edges.

That last part matters more than it sounds. Alongside position, points can carry colour values, intensity and normal vectors, which describe which way the surface faces at that spot. Scanning software stores these as attributes attached to each point. What none of them carry is any statement about how one point relates to the next.

How many points you end up with depends on the object and the resolution. A scan can produce anywhere from a few thousand points to a few billion. Our 3D scanning service captures to 0.04 mm using blue light and infrared, and we set the density against the part rather than maxing it out by default. A denser cloud isn't free. It costs processing time at every stage after it.

Why can't you just print a point cloud?

Because a printer needs a closed, watertight shape and a point cloud is a set of loose dots. Nothing in the file says which side is the inside of the part. The cloud has to become a mesh first, which means joining the points into triangles and closing the gaps.

That conversion has a name: surface reconstruction. It's needed because a point cloud contains no surfaces at all, only separate coordinates, and the software has to work out which points are real measurements and which are strays. Get that judgement wrong and you get lumps, holes and folded-over triangles.

There's a second problem. A scanner only records what it can see from where it's standing, so first-pass clouds arrive with empty patches where undercuts, deep bores and hidden faces should be. A mesh built straight over those gaps will bridge them with something smooth and wrong. Once the mesh is sound, industrial 3D printing from it is simple enough.

Cleaning and aligning, the unglamorous half

Most of the work on a point cloud happens before anyone models anything, and it's tedious.

Registration. Because one viewpoint can't see a whole object, a part gets scanned several times from different angles and the results are aligned into one cloud. That alignment is called registration, and it's where error creeps in. Academic testing on real scanned parts shows most registration methods start failing once scans carry more noise or need larger transformations. Small misalignments at this stage become double surfaces later.

Filtering. Real scans pick up points that aren't on the part. Dust, reflections, background objects and shiny or dark surfaces all produce stray readings. Cleaning uses methods like statistical outlier removal, which compares each point against its neighbours, and voxel grid filtering, which thins the data down.

Size. Scan data is heavy. Files reach tens or hundreds of gigabytes, and a full industrial digitisation campaign can run to billions of points and multi-terabyte volumes. Some of that gets thinned deliberately, which is a decision rather than a default. Thin it too far and you lose the small radius that mattered.

What can a point cloud become?

Three things, depending on what you need. A mesh, for viewing, 3D printing and rough replication. A surface model, for organic shapes that need smooth continuous curves. Or a parametric CAD model, where features and dimensions can be edited. The third costs the most and is the only one you can engineer with.

A mesh is enough when you want to look at the part, print a copy, or check it against a drawing. It's quick and it's often all a job needs.

A surface model suits sculpted and freeform geometry: handles, housings, anything styled rather than dimensioned. The surfaces follow the scan closely and stay smooth.

A parametric CAD model is the expensive one and the useful one. Features are defined, dimensions are editable, and the file can go into reverse engineering, redesign or production without starting again. Our product design team builds these, and from there a part can go straight into low volume manufacturing with no tooling.

Paying for the third when you needed the first is one of the more common ways to waste money on a scan. So is the reverse, six weeks later.

Which file format should you send us?

E57 if you have the choice. It's an open standard defined by ASTM E2807, so almost any software can read it, and it stores colour and scanner metadata alongside the coordinates. PLY, PTS, XYZ and OBJ all work too. Send what your scanner produced rather than converting it first.

The formats differ mainly in what they carry beyond position. XYZ is plain text with one point per line and three numbers as the minimum, which makes it readable by anything and heavy to move. Vendor formats hold more but tie the file to one software stack.

Every conversion risks dropping something, so keep your originals. If you're not sure what you've got, send it anyway and we'll open it.

What AI has changed, and what it hasn't

Point cloud processing has moved quickly. Mesh clean-up, splitting a scan into regions and recognising shapes are now automated to a high degree, and software can pick out planes, cylinders, holes and blends on its own. The tedious middle of the job is a lot faster than it was three years ago.

What hasn't changed is the decision at the end of it. The same analysis notes that engineers still choose whether a recognised feature should reflect the original design intent or the scanned condition of the part in front of them. That's the whole job in one sentence.

A worn part scanned faithfully gives you a worn part. Somebody has to decide whether that 11.94 mm hole was meant to be 12.00 mm, whether the flat face is meant to be flat or has sagged, and which surfaces are design and which are just how it was moulded. No amount of automation makes that call for you.

Two things to hold onto. A point cloud is measurement, and it needs cleaning, aligning and converting before it's any use to anyone. And the format you convert it into should follow the job, because a mesh and a parametric model cost very different amounts and do very different things.

If you've got a cloud sitting on a drive and no idea what to do with it, send it over. If you've got a part and no scan at all, send us that instead. Email sales@nexform.co.uk or use the contact form with a CAD file, a sketch or the part itself, and we'll come back with a quote and a recommendation on which output you need.