What it actually costs to process 167 photos

One real dataset, measured end to end: wall-clock time, peak memory, scratch space and output size. No estimates.

Most write-ups about photogrammetry cost either quote a licence price or hand-wave about "cloud credits". Neither tells you what one job consumes. So here is a single dataset, instrumented from first image to final download.

The run

167 nadir images of a quarry, GPS in the EXIF, no ground control points. Processed with OpenDroneMap on an 8 vCPU machine.

Where the 798 MB goes

The output is not evenly distributed. The textured mesh is over half of it, and it is usually the file people need least:

If you are paying for egress or storage, the mesh is the first thing to drop. Most measurement work needs the orthomosaic and the surface model and nothing else.

The number that surprises people

21 GiB of peak memory. Not 8, not 16. That is for 167 images — a small job by commercial standards. This is the single reason "just run it on your laptop" advice fails: a 16 GB machine will start the job, run for forty minutes, and die during meshing or orthophoto generation, after you have already spent the time.

Scratch space matters almost as much. This run wrote 10.25 GB of intermediate files, which works out to roughly 0.48 GB of final output per 100 images and several times that in temporary data. A nearly-full disk fails the same way a small machine does: late, and after the time is spent.

Scaling it in your head

At 18.5 s per image on 8 vCPU, a 300-image dataset is a little over 90 minutes of compute, assuming memory holds. Memory does not scale linearly with image count — it scales with the size of the reconstruction — so the honest answer for larger jobs is that you have to measure them, not extrapolate. We cap pilot datasets at 300 images for exactly this reason: it is the range we have actually measured.

The compute cost itself is small: a job like this is on the order of one to three dollars of cloud compute. The cost that matters is the fifty-one minutes, the 21 GiB, and the person who has to notice when a run fails.