In the case of point clouds - e.g. white points on black background - it seems pushing should be preferred, no? Would a pull in this case with inverted transformation introduce undesired blur on the transformed image or it depends on the chosen warp interpolation?
Yes, filling potential “holes” by using the pull mode would require some interpolation (and blur).
For visualizing a point cloud, the “push” mode could be better indeed.
I think it depends on who’s using it and for what purpose - you get less blur, but then you lost some positional information. It’s good to have the choice here!
In my case a point from the cloud is just re-positioned to a new coordinate keeping its intensity value so I think push is the way to go and I don’t see how that positional information is lost here. Care to elaborate please @garagecoder ?
Just a guess: the loss comes from the fact that pixel positions are discrete in an image, for example, if you want to rotate an image say 1°. In a way, a pull will be more informative visually, since wherever we see a blur, we know the position is in between pixels. We could potentially go backwards by comparing the weights of the pixels and what we know about the original image
Thanks @afre for the insight. I see how a blurred pixel can offer some information about the sub-pixel accuracy. In my case I think I would be happy to relocate a point to an integer location using whatever number rounding is appropriate - e.g. 10.5 goes to 11 and 10.49 goes to 10. Pushing should be sufficient for that.