Not sure whether upfront calibration will improve things. I once read you might be better off placing some well-structured object in your recording volume and calibrate the cameras from that.
Do you remember where you read this? I am very interested in setting up something like this to:
1) initially focus the cameras
2) using coded targets, establish world units as well as orientation
3) get a rock-solid solve on camera positions, which can then be re-loaded to other sets of images (particularly uncooperative smooth/featureless surfaces that don't align nicely otherwise)
I'm thinking a mannequin with noisy spraypaint applied, plus coded targets on the ground and a vertical bar, could produce this result. Curious if anyone has attempted this already?