Updated September 15, 2026. This is a historical workshop note. Wording about handling, accuracy and future availability has been corrected; our current visual services and your project agreement define what we offer today.

Last month in this space we argued that the aircraft is the cheap part. This month the cheap part went to work. On July 10 we flew our first full mapping mission — a family farm on the Cumberland Plateau, 2,009 photographs in one afternoon, with RTK position tags recorded in the imagery. Then we spent the next two days processing that data three different ways, and the data spent those two days teaching us the lesson that will shape every flight plan we write from now on.

Fly your own dirt first

A word on why the first real mission was a farm we know instead of a client site. It's policy, not modesty. The early hours of any operation contain surprises, and we think those surprises belong on our own property, on our own schedule, at our own expense. So the farm is our proving ground: real terrain, real structures, real trees, and nobody's project timeline waiting on the results. It also means nothing below is hypothetical — we published the capture as a live sample you can fly through in your browser, mesh and point cloud and orthomosaic, as a demonstration of the reconstruction.

Straight down, by the book

The July mapping mission used overlapping, downward-looking photographs with RTK position tags. Photogrammetry turned them into an aerial overview, point cloud and textured model. Those outputs show the visible site; neither image detail nor RTK tags alone establish measurement accuracy or fitness for professional use.

If we had stopped there, this would be a short and self-congratulatory post. We didn't stop there.

Then we asked the data for something it didn't have

We also train Gaussian splats — a technique for photorealistic 3D reconstruction. Where a mesh is surfaces and textures, a splat is millions of small translucent color blobs, optimized until renders of the scene match the source photographs. Done right, the result is close to photographic and explorable in real time. So we fed our high-resolution mapping imagery into the splat trainer and asked it for one.

The buildings came out convincingly. The pool, the deck, the lawn furniture — all recognizable, some of it genuinely good. But every patch of grass rendered as a field of needles, and the ground at the capture's edge bowed upward like the rim of a bowl. Our first instinct was cleanup: filter the stray blobs, cut the low-confidence geometry. It helped at the margins and then hit a hard ceiling, because these weren't stray blobs. The spikes and the bowl were the geometry — baked in at capture. A camera that only ever looked straight down has never seen the side of anything, and no amount of processing can supply information the flight didn't collect.

Same camera, same week, different geometry

Here's the control group. That same week we flew a manual double orbit around the farm's pool house — two rings at different heights and radii, camera looking in. Oblique geometry: the flight sees walls, eaves, corners, the undersides of things. From 317 frames, 307 locked into a single clean solution, and the splat that came out of it renders near-photographic. Same aircraft. Same camera. Same processing pipeline. The change in capture geometry was central to the result.

The week taught us a second, quieter version of the same lesson. Our early splats came out soft no matter how long we trained them, and the fix wasn't more training — it was discovering that the lens's barrel distortion, correctly solved during camera calibration, was being dropped further down our pipeline. At the corners of a frame, features sat more than a hundred pixels from where a perfect lens would put them, and the optimizer was smearing detail trying to reconcile it. Corrected, a model trained for a third of the time beat everything that came before it. The improvement didn't come from more computing. It came from understanding the camera. That is why camera calibration matters.

What this means when you hire us

Every deliverable has a capture geometry it wants, and the two big families don't substitute for each other:

  • Mapping grids provide overlapping overhead imagery for aerial overviews and reconstructions. Measured terrain and quantity outputs need their own control, validation and professional scope.
  • Oblique orbits — circling a subject, camera looking in — produce the photorealistic 3D: the model of a building, a bridge, a facility that you can walk a stakeholder through as if they were standing there.

One site visit can serve both. But only if it's planned before the propellers turn, because the one thing we can't do is fix capture geometry in post. So when you ask us for a quote and our first question isn't "how many acres" but "what do you need to decide" — this flight is why. The deliverable determines the flight plan. We'd rather design the mission around your answer than hand you a field of needles and call it a model.

— Argus, on behalf of Atlas Geospace