New price target $2, page-38

Photogrammetry has come a long long way. Main bit of magic here is an algorithm called structure-from-motion (SfM).

Take a bunch of photos, calculate (or know) the pose of each photo, you can then compute the location of matching pixels in 3-dimensional space.

Each matched point .. billions of points .. a point cloud .. gets joined with its neighbours to form a triangular mesh. As part of the reconstruction, the photographs are then matched to the polygon mesh to provide a texture.

I still love watching the tiles (of mesh data) as they come out, but I am simple ;-p

Plenty of packages implement this - ContextCapture, Pix4D, Metashape, Reality Capture, Skyline Photomesh .. etc etc. We use Context. Last I heard Nearmap still uses Context. Also the reason Spookfish cameras don't work - they "photos" are more like how a fax machine or satellite captures imagery via a swath - so the "image" parameters aren't truly known and the reconstruction struggles.

Aerometrex play in this space .. but the compute resources to do this for entire cities .. every couple of months .. is absurd. Nearmap generate the data at about 12cm last I looked - but with 4cm imagery and enough compute .. you get the idea.

For a full city .. note the Bentley logo .. who make ContextCapture.

Below is the result of 7,500 drone photos from Hayman Island post-cyclone .. run through ContextCapture .. with survey-grade control ... you get this output - a full 3D mesh .. red circle is zoomed in below .. this was nadir + obliques for about 1.5cm GSD - so each pixel from the photo was about 1.5cm across on the ground. 3D accuracy is circa 50mm. We have run meshes down to the mm GSD range using handheld cameras - final sample below.

This is still dumb data. Throw an AI/machine learning workload against it and things really start to get interesting .. this is still a point cloud underneath - see final image of the mesh - which was generated from a coloured terrestrial laser scan.

For what it is worth, we used the point cloud from this model back in Cyclone and subsequently Revit to model roof data in some cases. We actually laser scanned the place as well but there was the odd roof we couldn't get on.



Zoom of red circle area - this 3D mesh is from drone images stitched together in 3D space.



Different model - this was just coloured points from terrestrial laser scan .. then meshed .. with mesh visible .. if you look at the bricks you can sort of make out the points from the initial point cloud - but the point cloud would have been circa 2 billion points - one day's collection.