Making Sense of Errors.

[diggingforbones]

Hello all, thanks for stopping by. This is quite tricky, at least for me, to get my head around.

I've created a few models with UAV photogrammetry, utilising cheap direct-georeferencing. I know that it's not very accurate.

My question is, how does Photoscan calculate the error of camera positions without having a known, fixed, georeferenced point to compare to?

For example: the Total Error of these camera positions is 0.949: http://imgur.com/a/M7Jnb which is massive!
Yet the Total Error of my scale bars is fairly low (calculated by inputting values that were manually measured in the field).

If anybody knows how the error of camera locations is calculated then could they please explain it to me? My thinking was that by lowering the camera location error (by using RTK or GCP referencing), you will generally end up with better scale measurements.
Thank you.

PS: Extra bonus discussion point! Does good photography then negate the effect of bad georeferencing?

[Jeremiah_ROWE]

Agisoft "error values" are in regard to the correlation of the camera orientation and tie-point position value, they are not accuracy values! They are the average errors in regards to the overall fit of the model relative to its contents.

If you want to actually check accuracy you need to use high-precision survey grade checkpoints to compare against the surface you generate.

Your data's final accuracy can be no better than the accuracy of your input geo-data.

Without using survey measured GCPs and checkpoint data, there is no way to fully determine your own accuracy.

[diggingforbones]

Hello this makes a lot of sense! Thanks for taking the time to explain that to me.

To clarify, these errors values are a measure of how the camera positions have been shifted to make nice tie points (or to match up the features nicely)?

May I also ask where you learned this information from? I didn't get much from the manual or the rest of the forum. I believe you as I also asked on reddit and the answers are essentially the same. It's just interesting to me that it's not expressly published anywhere!

Thanks again

[stihl]

That is correct. The error shown is a shift that Photoscan had to implement in order to achieve a good set of tie-points according to the results of the image matching. These should then be further refined by the use of survey grade ground control points.

The GPS of your DJI  drone should not be considered as accurate. Furthermore, the altitude is usually wrong because its zero value is usually the place where it took off. I.E. it's a 'height above ground level' (AGL) value and not one that's referenced to the WGS84 ellipsoid which is the case for the X and Y coordinates.

Information about this subject can be learned by taking photogrammetry and or surveying courses or studies.