Point cloud stitching without reference markers?

[[Yi...]]

Hi everyone,

I’m wondering whether ATOS/ZEISS Inspect can align multiple scans without reference markers, using point cloud geometry (e.g. ICP) only.

If so, how does the performance of this method compare to that of marker-based alignment?

Any official guidance or reference papers on this would be much appreciated.

Thanks in advance!

Yixiang

[[Ma...]]

You can use bestfit for aligning meshes from scan series, but you will loose accuracy.

Also you have to consider shape of your part. If it's symmetrical, then you will struggle with aligning - SW won't know which side is which.
Also wrong placed ref. points won't help, because you can either use bestfit or common ref. points.
On bestfit you can help SW by clicking on two meshes to tell that those points are close to each other, but you have to learn how to use it.

I've performed scans of two dies / each die had it's own program / i've performed geometrical alignment ( even without cad model ) and exported mesh in that alignment.
In another program i've placed those two meshes into program as measured data ( one was in clipboard ), did manual translation for one of meshes and used merging.
This way i was able to obtain a forged part from parts ( negative shape ) - after inverting faces of the mesh i could make measurements.

So many thing are possible - it's just learning how to use it properly.

[[Yi...]]

Please sign in to view this quote.

Hi Martin,

Thanks a lot for your explanation!
Based on your observations, how much accuracy do you think is typically lost when using bestfit?

Actually, my question is more like this:
If I understand correctly, the point cloud density in ATOS should correspond to the camera’s pixel density. So, in an ideal scenario, is there any difference in accuracy when calculating the transformation matrix from measuring centroid of reference markers compared to calculating it from overlapped point clouds using bestfit?

Best regards,
Yixiang

[[Ma...]]

I did not do that study, but SW is telling us some calculated error about 0,01...mm, but to be honest we are scanning sandblasted forged parts, mostly without spray, so our mismatch can be different than on machined part.
This error is only deviation of selected points used in bestfit - you can still end up with mismatched part ( like inverted, rotated ) and having a little bestfit error.

What to have in mind is that bestfit will use only overlaping faces from both scans.
For flat parts i would recommend scanning like in vice to obtain more like 75% of that part. Then next serie would be missing part - that would help in bestfitting or to have at least 3 ref. points on body.

There is rule, that you have to have at least 3 ref. points from previous scans in current scan serie.
So by example with using vise - i would place at least 4 ref. point on a part to be visible on both scan serie - try to keep them further away from themselves to help SW define volume.
If you imagine you place 3 ref. points like long line, then SW will end up with long triangle with tiny height - that is like small datum in Calypso. Using 4 points to create something like plane in 3D it will help a lot to fit by ref. points.

[[No...]]

Generally, you CAN stitch scans by geometry only. You MAY lose some accuracy, but this depends a lot on the part and how much overlapping you can provide to the software. You should always aim to have as much overlap as possible, ideally there should be distinct geometry in all three space dimensions, so the fitting algorithm has something to work with.

For example, if you scan the top and the bottom half of a plain cylinder without additional features, you may have enough overlap perpendicular to its axis (2 dimensions), but along the axis (3d dimension) there's almost nothing to use for fitting, so the result will probably be bad if it works at all. If however the cylinder has one or more perpendicular bores at distinct positions, that you can include in the scan, fitting can also be performed well in the third dimension and you may get good results.

Edited 4 hours ago