I'm going to answer this as best as I can, although I've also found dealing with tforms at different scales to be quite tricky. tforms can be difficult to work with because they rotate and project around the coordinate 0,0 and not the image center, so it's difficult to pull them apart and understand each underlying transformation.
-For instance, what are the units (if any) of MaximumStepLength?
This is a parameter passed to the optimizer registration.optimizer.RegularStepGradientDescent. It doesn't have units because it isn't related to the image information and is an optimizer parameter: https://www.mathworks.com/help/images/ref/registration.optimizer.regularstepgradientdescent.html
-If I pass the spatial reference to the function, the tform is supposed to be in the same units as the reference. Is that correct? And if so, how can I pass an 'InitialTransformation' parameter that was obtained using a different scale?
All of the elements of a geometric transfomation matrix are scale-indepdent with the exception of translation. So in order to rescale tforms you need to do tform.T(3,1:2) = tform.T(3,1:2)*ScaleFactor: https://www.mathworks.com/help/images/matrix-representation-of-geometric-transformations.html
To explain what I am doing, I have two sets of slices from a microscope scan. The number of slices is much smaller than the XY resolution (1040*1392 and 60 slices) so I am implementing my own pyramid algorithm ,in which X and Y resolutions are halved at each level but Z (n of slices) remains constant. I want to find the best transformation at each level and pass it to the next one but I suspect I will need to rescale the tform before passing it.
Given image of size 1040x1392 with a corresponding tform aligned at half resolution, your code will look like: tform.T(3,1:2) = tform.T(3,1:2)*2; registeredImage = imwarp(image,tform,'nearest','OutputView',imref2d(size(image)));
