Three square kilometers, six weeks, two tripod scanners, one handheld, and a client who wanted Revit models and 3ds Max scenes at the end. LiDAR and photogrammetry were the visible part. The pipeline was the part that decided whether the six weeks ended in a delivery or a folder of unregistered scans. This is the shape of that pipeline, including the parts that failed.
Start with tiles, not with scans
We divided the district into 250-meter tiles before anyone set up a tripod. Each tile had an ID, a footprint, and an expected scan count based on street frontage. The scan index was a CSV with easting, northing, and a scan ID, which made coverage a Python problem instead of an argument.
That script ran every evening during the field phase. If a tile showed up thin two days in a row, the next morning went there first. It sounds trivial. It saved us from discovering a coverage hole in week five, when re-scanning means permits all over again.
Two field habits made the manifest trustworthy. Every scan ID was written on a card and photographed at the setup, so a file name could be traced back to a corner. And every day's folder got a checksum line in the index, which caught three card copies that had silently truncated.
The manifest also carried the non-scanned ring. For those tiles we generated context with GenAnywhere by selecting the area on the map, and stored the result as a separate layer with its own flag. Keeping the two sources apart in the manifest was the difference between a usable model and a mess nobody could audit.
Registration order matters more than registration settings
Autodesk ReCap will happily align a hundred setups into a point cloud that looks convincing and is wrong. We registered in groups of twenty, checked each group against total-station control, then merged. Group-level residuals were recorded in a CSV next to the project, not buried in software.
The final checkpoint RMSE came out at 17 mm, with one corner at 29 mm that we traced to a control point set on a loose paving stone. We reset it and re-ran the merge. Boring, repeatable, and the only reason I could sign the delivery.
ReCap has a habit of making a bad merge look good. The only defense is a checkpoint set that covers the edges of the project, not just the middle. Ours included six points on the far side of the river, which is where the worst residual showed up.
Photogrammetry as a second opinion
The pole camera images went through photogrammetry for facade texture and as a sanity check on geometry. When a laser surface and a photo surface disagreed by more than 3 cm, we looked. Most disagreements were texture projection, a few were actual scan noise from reflective glass, and one was a wall that had moved since the control was set. The photo sets were named by tile and height, which made the projection step a loop instead of an afternoon of clicking.
From cloud to Revit to 3ds Max
The core cloud was decimated for Revit links and kept at full resolution in archives. Modeling was done against sections and orthophotos, not against a spinning cloud, because people model better against a flat slice. The 3ds Max scenes used a lighter mesh plus projected texture, exported as glTF for the client's web preview. That was scan to BIM with a rendering detour.
The web preview was the surprise requirement and the easiest deliverable. Tiling the export by the same 250-meter grid meant the viewer could stream it, and the map-based context layer sat underneath at low resolution. Two inputs, one coordinate system, no drama.
What I would automate next
Coverage checks ran nightly. Checkpoint reporting was manual. That is backward. The next pipeline gets the control sheet generated from the same manifest, so a tile cannot be marked done without its checkpoint rows filled in. A historic district scan is mostly logistics, and logistics is mostly bookkeeping with tripods.
The scan cards became a spreadsheet by week three, which is late. Next time the card is a QR code printed from the manifest and scanned into the index at the setup. Same information, one less transcription, and one fewer argument about which scan covered the corner by the pharmacy.


Top comments (0)