SmartFilter reads a 3D capture and separates it into the things it is made of — floor, ceiling, walls, pipe, ductwork, structure, clutter. You can then show only what you care about, isolate a level, or cut the model open.
It works on the 3D surfaces: the 3D map and drone captures. It does nothing in first-person mode.
Three tabs
Segments
A list of classes, each with the percentage of the model it accounts for. Turning a class off hides it. Classes seen in the build include Roof, Floor, Ceiling, Wall, Beam, Column, Building, Dirt Ground, Debris/Clutter, Construction Materials, Equipment and Unlabeled.
The full set is twenty labels, and it is the same twenty everywhere — one flat list, no families or groups above it. If your capture shows fewer, none are missing. Labels that do not occur in the capture, or that belong to levels you have not loaded, are shown disabled and marked Not in loaded levels rather than dropped from the list. So two captures in the same project can offer different active classes and still be running the same taxonomy — load more levels and the greyed-out entries come back.
Show unsegmented in true color keeps everything the classifier could not label in its original colours instead of hiding it, so you can see what was missed.
Multi-Level
Controls which levels are visible in the dollhouse at once, so a multi-storey model can be shown one level at a time or several together.
Clipping
Cuts the model with planes so you can see inside it. Relative Range resets the cut back to the full extent.
What it is called
On CupixWorks 4.0 this was Semantic Segmentation. It is SmartFilter at 5.0, and the engine underneath is the same one.
What SmartFilter does not label
SmartFilter labels what a thing is, never what condition it is in. The twenty classes are building elements — floor, wall, duct, column. There is no surface-quality family: nothing here marks corrosion, water damage, concrete defects, spills or cracks, and there is no confidence score on a class either. Confirmed by the product owner, 2026-09-15.
It is worth saying because the opposite is easy to assume from the name. If you came here looking for defect detection, it is not in this panel — and the classification itself is automatic, so it is worth a look before you rely on a percentage.