A similar issue (and possible solution):
I want to add images to an existing alignment, I don't mind having to re-align all images, but the images that are already aligned should stay in approximately the same position.
Concretely the situation is:
- I've just processed a dataset of ca 15 000 video frames of an underwater shipwreck site. Because of bad visibility, bad lighting and bad overlap there has been a lot of manual work involved in getting all the images to align (a lot of "reset camera alignment, align selected cameras, manual marker placement and optimisation"), so I've worked on it for various days in order to get this far.
- Now I've received 400 additional, good images of the site from my client. The new photos probably won't contribute much to the geometry but could be very good for texturing.
=> I want to align these 400 images to the remaining 15 000 video frames without changing the position of the original images (too much).
Would this be the correct workflow?
- In the PhotoScan document with 15 000 aligned video frames, save the camera calibration of each camera group from the Camera Calibration menu.
- In the PhotoScan document with 15 000 aligned video frames, go to Reference pane, click "export", choose ".txt" format (PhotoScan says: "can't export reference data when I choose .xml)
- export the "cameras" making sure "save location" and "save rotation" are checked. In my case the estimated values will be saved, since I don't have ground control coordinates.
- make a new PhotoScan document, import the 15 000 image frames previously aligned.
- in the Camera Calibration menu, import the previous camera calibration of each group and check "fix calibration"
- in the reference pane, click "import", choose the .txt file containing my old reference data, and set XYZ and Yaw, Pitch, Roll column to match the column containing the estimated values from my previous alignment. If I now switch on cameras in 3D view, a blue ball appears for each camera location (shouldn't this be a blue square with a line, to show that the camera rotation is also known?)
- under Reference Settings adjust Camera Accuracy to something small like 0.05m => ie the cameras can only move 5cm from their current location? (how does this work when the alignment is not scaled yet?)
- add the 400 new images to the same PhotoScan document
- align all 15 400 images together. My settings would be: medium accuracy, generic preselection on, reference preselection on (so it uses the previous alignment), key point limit 40 000, tie point limit 0, adaptive camera model fitting checked (unless someone has suggestion as to why another setting would be better).
Hopefully by importing the Reference data and setting a strict camera accuracy PhotoScan will keep the old positions (more or less) and have an easier time aligning the images relative to each other (when compared to the very manual first time alignment)? I would appreciate any input from the community and developers!
Cheers,
Thomas