Measurement Accuracy of 3D Models?

DroneDeploy has published estimated RELATIVE measurement accuracy (horizontal plane) for maps created without Ground Control Points based on GSD. I realize the accuracy of 3D models largely depends on the number of photos taken from different altitudes and gimbal angles and the resolution of the surface area of the images - effectively the “X/Y/Z plane GSDs.” Using an “optimal” set of flight plans to build a 3D model of a structure with effective distances from the surfaces of 20 feet using a 20MP camera, such as a communications tower, has anyone, including DroneDeploy even have determined an estimate on the relative accuracy of measurements of a 3D model in X, Y, and Z axis?

With good control points or RTK we are seeing 0.10ft/100ft horizontally and 0.15-0.20ft vertically. There are so many details that could skew these numbers which is probably why nobody publishes it.

My exact thoughts. Out of curiosity, I contacted DD Chat Support, and as expected this is WAY beyond their level of expertise to answer. They said they’d pass it along and get a “second opinion”. They referred me to the RMSE stats of the project which I pointed out are ABSOLUTE position RMSE!! They of course don’t publish relative measurement RMSE. If I recall the RMSE on relative measurements can grow as the measurement distance grows. We tested on a football field (since it has so many painted lines) and a tape measure. At 0.75 inch GSD we were seeing up to +/- 4 inch error depending on how far we measured. That’s a lot more that the BS 1-3X GSD you see published all over the Internet including by DD. And that’s just 2D.

1 Like

How were you ground measurements obtained? Obviously the error is going to increase as it is a percentage based value. You would see the same thing with an RTK drone but not near as bad. Don’t forget we are using grid on the ground and the drone is geographic which means that there is also error in the transformation. I just measured a football field a week or so ago and horizontal was 0.04ft on 100yds. GSD was 0.5in/px and the measurements came from verification of their striping with an Emlid RX. You didn’t state your measurement length.

End zone line to 20 yard line and midfield line with a tape measure. So basically 3 measurements, as we measured midfield to other endzone and added the two and compared to measurements using DD and Metashape formerly photoscan. It was a two-fold experiment to compare processing software accuracy and actual accuracy relative to GSD. Phantom4Pro. No GCPs. I don’t remember the overlap used, this was 4+ years ago and DD has updated its photogrammetry software significantly since. I think endzone to endzone which is what 120 yards was 5 or 6 inches off. You probably have to factor a +/- 1-inch “fudge factor” in as you’re not going to place the tape measure at the exact precise spots as the measurement tool endpoints in either software unless you’re doing like 0.2 inch GSD maybe.

That’s when I disregarded all the published report that horizontal plane accuracy was 1-3X GSD. That’s optimistic without RTK or GCPs. Also did a couple years later a very small area at like 0.2 inch GSD 50ft x 50ft area, flat pavement, painted markers and got out the tape measure. Can’t remember the figure but it was unexpectedly inaccurate based on the expected accuracy at that GSD and high overlap.

yeah, tapes are fun. Trying to keep a 200ft’r square and level with the right tension. It isn’t easy when you’re trying to maintain inches. Without GCPs it’s not super surprising, but I would expect a little better at 360ft, more like 2-3in. How relative were the DD and Metashape results?

Too long ago for me to remember. It must have not been a big difference as we would have used Metashape a lot more for simple mapping work since then if it had superior photogrammetry accuracy.