Chunk transform changing after alignPhotos sometimes

[Rafes]

Hi,

(I'm not sure how many variables to explain here, so please let me know if there are any ideas that I can follow up on for more information.)

The basic setup:

• I have a lot of cameras in a chunk. Some are already aligned by importing cameras .xml.
• The cameras that are not aligned all have very accurate reference data for the location
• I used the python API to matchPhotos

I reset the transform for this chunk and duplicated it twice before continuing, we'll call them "A" and "B"

In chunk A...
I run "Workflow > Align Photos..." 
The chunk transform changes!

In Chunk B...
I clear all reference data
I run "Workflow > Align Photos..." 
The chunk transform does not change!

I then reset chunk B and it essentially aligned back to what chunk A looks like (not exactly because they are 2 different solves, but obviously correct anyway.)


My question is why is the chunk transform changing only when reference data is being used?

[Alexey Pasumansky]

Hello Rafes,

If geographic/projected coordinate system is used in the chunk, then chunk.transform.matrix defines the transformation from the internal coordinate system used in the chunk to geocentric coordinate system.

I would expect to have different chunk.transform matrix when alignment is reset and performed from scratch. But does it bring any inconvenience on your side?

[Rafes]

But does it bring any inconvenience on your side?

It feels non-deterministic If we remove a camera, or do more solves of the content over time, where cameras can shift or fail to align for some reason, then the resulting reconstruction shifts in "world space". It is possible I am misunderstanding a concept or misusing the resulting data. I will send you a DM

[Rafes]

Note the cameras no longer match the reference data when this happens. The whole chunk changes and if I reset the camera, the reference data is in the original place/space. It isn't clear to us how reference data is used during alignment, especially in this case. I would really love to have Metashape obey the reference data as a hard constraint (or at least have deterministic control). If I say a camera location is accurate to 1mm, it really should not be moved more than 1mm.