Comparing two models from Agisoft

[benjamins]

Heeye!

I got uav Pictures of a river basin in mongolia from 2015 and 2016 and my task is to compare the decrease and increase because of Erosion. So i processed them in Agisoft Photoscan (georeferenced, calibrated) and build a 3d model. I exported both models in a '.ply' file and loaded it into cloudcompare. So now i simply registrated them with 'match bounding-box centre' followed by the more accurate 'ICP registration' followed by the 'compute cloud/mesh distance'.
Actually my Problem ist that the results dont look really expected/realistic and i dont know why ..maybe the Registration didnt work well cause it Looks really strange after icp Registration (the red colour actually show the 2016 model and the other one is 2015). The strange thing is that after i compute the distances (before pressing the compute button) the Approximation looks pretty logical and good but after i compute it Looks so different ..
mhh maybe i missed a big step or something ? like do i have to scale the model if its already georeferenced ?

sorry for my bad english :/ hope u understand me

[benjamins]

Maybe it's not a problem with cloudcompare ... i thought about another reason. Is it possible that the heigths in the elevation modell are distortioned because of a bad marker placement for georeferencing in photoscan ? The pictures below are showing the marker placements that i got from the reference year 2015 and they are actually not wide spreaded (marker placements from 2016 are actually pretty good)... could this be a reason ?

[daniel]

Indeed, there are a lot of potential issues:
- the fact that the GCPs are not widespread (have you tried to compare the entities without performing the ICP registration by the way?)
- if you use ICP, then make sure to use a relatively low 'overlap' ratio (e.g. 50%, to take into account the fact that the overlap is not the same AND there must have been some structural modifications
- I don't see the result of the distances computation... can you explain a little bit more what doesn't seem right?

Don't hesitate to send me the data if you me to help you more precisely. And have you looked at the M3C2 plugin? It's perfect to do this kind of comparison.

[benjamins]

Hey Daniel!

thanks for your reply. Below are the results of the distance computation (scale is in meters). In my opinion it doesn't seem right because in 1 year it's not possible to change that much ... even that almost everything changes is strange. Should be more area without change i guess and i thought that other spots would be more significant like the river sites/banks. About the second fact i am not sure but the first one is really strange ..

The third picture is the result after icp registration. You can already see the big difference in the right corner, that's why i guessed that the registration wasn't right or the model is bad (cause of gcp).
By the way thanks about the tip with the M3C2 plugin. It seems to be perfect for my situation but i dont know if it makes sense with such a strange registration ?!

If u still think it would make sense to have the data by your own i would like to send it to you via cloud link or something ?!

[daniel]

Looking at the shape and distribution of the deformations, where you have small differences in the middle (where the GCPs lie?), and wide differences on the border, I fear that the issue might be more with the photogrammetric process.