Stereo Cameratracking workflow (PFtrack2011)

[Dashinuke]

Hey people,

I'm tracking a stereo shot and I run into a few problems. The shot is a crane shot, and the first question is; Is it actually possible to solve a Nodal-pan/rotation in stereo ? Since with a nodal solve, the pointcloud doesn't get any actual depth (and therefore, test-object always sticks, no matter in which depth they are). However, you need a decent 3d scene to place the object in, to get the right stereoscopic depth between the two eyes?

Second, I notice that in the 2d-track node you can set-up the consistency of the tracks (free move / rotation, etc) but I cant decide in the solve-camera stage how I want the camera to solve? if I want to solve it differently, I need to do 2d-tracking all over again ?

Third, is it wise to stereo-solve two camera's, or is it better to solve just the hero eye, and create a parented second eye in a 3D app ? Is there a general accepted workflow ?

And finally, if it is actually wise to stereo solve it, you do the ocula-related corrections before the cameratrack, or save it till the end of the workflow ?

Thanks in advance!

[stefan]

Hey Dashi,

it is possible to solve rotation only shots in stereo. But... in depends on your stereocams io and the scene you have.
I worked on a diploma thesis where I tried to figure out when it works.. turned out, that shots with a deviation according to your closest and farthest point of more than 3 % can give you quite good results, whereas a deviation of 1% like you would have on features wont give you accurat results on just a rotationcam. But if its possible you can still take some helper frames and the second cam will in most times improve your results( if its synced and correctly aligned of course...)
When you cant get any good results, you can still adjust the second camera io in your 3D or comp Programm to match the distance of the object you want the scene to enhance with. I did that in Nuke with a rotation shot where I stabilised it, then added my 3d Object und then adjusted a second cam to fit to my second view which worked quite nice. Then remove the stabilise transformation and set the output of the scanlinerenderer to matchmove. so thats quite a workaround.. but fast.

Back to your third question... a stereosolve improves your solution in most of the times, depending on sync, alignment, occluding objects and all those obvious things. Your overall error can increase in fact... but due to the second view (which distance should or at least could be known) the triangulation of the object point will be more robust.

Ocula alignment should be done after the solve. Depending on your vertical alignment method you shift the pixels like you do with a cornerpin and that would be a funny sensor that gives you such results....

Hope you could get some informations out of that.

cheers Stefan