Hi all,
I’m currently working on building an orthomosaic of a 4,7ha big agricultural field (wheat) with Agisoft Metashape 1.6.2.
The images were taken with an XT2 (12MP, 8mm focal length, 9.5mm sensor width) at an altitude of
70m, leading to a GSD of about 2.5cm. In total 425 images were taken with a side and frontal overlap
of 80%. Image quality of each image – according to Metashape – is bigger than 0.8. The image
reference was deleted and the reference of the chunk was set to the right one (EPSG: 31256) in order
to avoid any complications.
Throughout the generation process I came up with 3 different approaches to do so. I will try to lead you through my workflow:
Initially (Approach 1), after aligning the images (accuracy: high, key point limit: 40.000, tie point limit:
4.000, generic preselection) only about 105 images were actually aligned. But using the “align
selected cameras”, when right-clicking the previously non-aligned cameras, aligned all cameras –
resulting in a tie point cloud of about 340.000 points.
The RMS reprojection error of the point cloud is at 0.275081 (1.53462 pix), max reprojection error at
1.40366 (46.6136 pix), mean key point size at 5.34687 pix, and the average tie point multiplicity at
2.64979. The maximum amount of valid matches ranges from 299 – 1300, depending on the image.
I then set 12 GCP’s leading to an RMSE of 0.71m – which is completely fine for this field, since the
orthomosaic only needs an approximate reference. The tie point cloud looks good and the position
and rotation of the field also seem to fit but it seems like Metashape thinks, that the images were
taken from below the ground (“Screenshot CameraPosition”).
If I then want to calculate the dense point cloud, I get the “Zero resolution” Error.
Then I went in the Forum and searched for solutions but could not find any that worked for me. I
already tested changing the alignment accuracy (to medium and highest), unchecking
generic preselection and changing the tie point limit (10.000 & 40.000) (Approach 2).
When increasing the tie point limit to 10.000 and deactivating generic preselection (no other
alignment parameters were changed), I got a tie point cloud with 465.000 points, RMS reprojection
error of 0.215251 (1.4669 pix), Max reprojection error of 0.719796 (63.9647 pix), mean key point size
of 5.10988 pix, and an average tie point multiplicity of 2.75229. I then set 5 GCP’s to just see if it
works.
With that tie point cloud I am able to calculate the dense point cloud but it looks like “Screenshot
DenseCloud”, which obviously is rubbish. Nevertheless, I’m able to calculate a DEM and an
Orthomosaic ("Screenshot OM"). On first sight it looks good (except for the edges), but if you zoom in
you can see, that for example the lanes of the tractor are offset (“Screenshot OM-Zoom”).
I then tried to repeat the second approach (Approach 3) with an even higher tie point limit (40.000) but again get the “Zero resolution” error, when trying to calculate the Dense Point Cloud.
Do you have any help?
Am I using an insufficient amount of photos?
Do you think setting more GCP’s on Approach 2 could improve the Orthomosaic?
The storyline of my post might be confusing now – and I’m sorry for that – but I just wanted to inform you about the things I already tried.
I used this type of workflow and parameters in a couple of other projects and never had any problem
like that.
Thank you very much!
Matthias