[jgord]

Fitting lines to 'xray' scans of buildings - turning pixels into vector art.

Lets say you have a complex industrial plant, or datacenter you want to upgrade.

You scan it with lidar and get a pointcloud and 360 panorama images. This gives you a large dataset, but what you really want is a floorplan, a lite CAD plan showing the racks, cable trays etc.

You take the scan, slice the pointcloud and make an ortho image .. it really looks like an xray of a building from the top down.

Then someone has to manually trace that in CAD to make a useful 3D model they can use for designing the upgrade.

So Im automating the boring manual part - turning the xray plan pixels into vector polylines, using machine learning.

One of our clients scanned their datacenter, and we generated a floorplan that shows all the rack box positions, cable trays, pipes etc.

Other examples : drawing the weld lines of patches in steel storage tanks, drawing in all the steel girder beams in a scan of an old railway bridge, or the windows, doors, ceiling pipes of a commercial realestate refurb.

gord at quato.xyz

As part of this work, were looking at running our custom machine learning kernel on multi-core x86 CPUs.

[haritha-j]

Interesting, so you process the point cloud purely in 2D slices? do things like vertical piping cause any issues? or is 2D enough? I'm just finishing up a PhD in modelling industrial spaces from point clouds so I find this space fascinating.

[jgord]

imo, 2D is maybe 70% of the problem to solve. We build and we think about buildings in 2D, possibly because its simpler/cheaper, possibly because of gravity corraling us into spaces with flat levels.

Computationally 2D is simpler, less dof, also conceptually. I decided to bring the scope in to get some wins on the board and early revenue.

I did do a lot of prior work in pure 3D, detecting walls and pipes :

  https://youtu.be/-o58qe8egS4
  https://youtu.be/8fjHNDGKeu4

ans also have some demos of fitting textured triangles to make lite 3D web browser models .. a while before gaussian splats came on the scene.

You are quite right, pipes do bend and jump thru 3D in complicated ways - and so flat slices only pick up part of their journey.

My rationale is most people want an accurate floorplan first, and quickly, and this saves time and captures much of whats needed.

Obviously the same 2D technique can be used in sequence - floorplan first, then get a slice of a wall, and get vector lines for all the door and window openings, or the stairs using the same process [ pointcloud slab slice -> ortho xray image -> vectors -> CAD ]

Feel free to contact via email