Tweet
I have three synchronized videos taken from three different cameras with unknown transformations between them. The object of interest is long, and I want to model this problem correctly in 3D. However, I am facing some difficulties, particularly in how to model the sequential approach between groups of cameras with an unknown fixed transformation. How should I proceed?
-
-
Hi,
as long as there's enough overlap, you should be able to simply extract all frames from the 3 videos and run the reconstruction. If you know the relative distances/rotations, you can define the camera groups in the wizard.
Even if not, you might still want to manually define three camera calibration in the SfM new project wizard, and assign each calibration to each group. Of course you don't need to do a calibration for each device and you can let zephyr to the work, but it helps in this case (make sure to allow zephyr to adjust the parameters) since you know there are 3 different devices. Being videos, you won't have exif data to rely on so this manual step may be helpful.
If the object is long it's possible the final result may suffer from drifting errors if there isn't enough coverage - ideally this won't happen, but if it does, use GCPs (available in the full version of zephyr) to control and/or drive the reconstruction. Make sure your capture isn't just a long series of videos but try to add more parallax / "context" of the object when shooting. -
Thanks Andrea!
Actually, the overlap between frames is quite good, but not between cameras, so that's the main reason why reconstructions are failing so often.
Main reason is cameras relative positions but also lack of some mutual features. One could expect that fixing the transformations between cameras, the system could use some parts of the (long) object to find transformations between cameras, but not in all frames.Comment
Comment