search

OpenDroneMap / ODM

A command line toolkit to generate maps, point clouds, 3D models and DEMs from drone, balloon or kite images. 📷
https://opendronemap.org
GNU Affero General Public License v3.0
4.68k stars 1.07k forks source link

Feature request / Proposal -- increase default GPS-Accuracy to 2̶m̶ 3m #1748

Closed smathermather closed 3 months ago

smathermather commented 4 months ago

For a couple years now, I've been of the mind that the default GPS-Accuracy under-constrains the model using GPS values, resulting in calibration issues which presents as excessive doming in the elevation data produced. Based on processing dozens of datasets on the community forum, the accuracy (or in this case, really a measure of consistency) for individual flights (rather than aggregate datasets) can be safely tuned to 0.2m / 2 decimeters.

Doming in dataset from Monrovia, Liberia

image

data credit: Community mapping in Monrovia, Liberia, with Slum Dwellers International. Residents are creating drone imagery, ditigitizing buildings, and conducting enumeration surveys using OpenDataKit and the HOT Field Mapping Tasking Manger.

For aggregate (multi-flight) datasets where there can be GPS consistency differences over time, it's safer to set this value to a larger value such as 2m.

I'm currently testing the switch to 2m with roughly a dozen community datasets shared with me over the last few years: ezgif-6-1a59c8ad6c

Thoughts? Any good counter examples? I have yet to break a dataset by reducing to 2m, but I could be missing salient examples.

smathermather commented 4 months ago

Community conversation on this here: https://community.opendronemap.org/t/default-gps-accuracy/19510

pierotofy commented 4 months ago

I think this would warrant a decent amount of datasets testing; over-constraining the BA problem this way might make the program more sensitive to GPS measurement errors, causing failures when such measurement errors are present.

I don't know how often such errors happen in practice. Changing the default could be OK if it's beneficial for most datasets (and doesn't substantially raise the failures).

smathermather commented 3 months ago

In the dozen sets I tried, not a single failure at 2m. Sounds like @Saijin-Naib uses 3 and has yet to see a failure.

Saijin-Naib commented 3 months ago

The only time I need to relax it is when using ground-level mobile phone collects with my TeraCube, but even the default 10m isnt nearly coarse enough. 30m is normally needed

smathermather commented 3 months ago

https://github.com/OpenDroneMap/ODM/pull/1749

smathermather commented 3 months ago

Edited to 3m on @Saijin-Naib and Mike_F's recommendation.