Alternating block solve

[minnerbe]

Changes

I added a client that does a couple of rounds of the distributed block solver with alternating layouts to ensure that every tile is in the "interior" of a block at least once. More specifically, the algorithm looks like this:

1. solve blocks with overlap
2. instead of interpolating transforms on overlapping regions, choose transform randomly from all candidates
3. save blocks to database with coherent naming scheme (_runN)
4. change parameters:
   • change stack to pull data from to target stack of previous iteration
   • shift block layout by half a block (before expanding, i.e., by 1/4 of an expanded block)
5. go to 1.

Test

I tested this on a small toy stack (1 mFOV):

• For comparison: here is a stack that was solved globally
• This stack was solved using a 3x3 block layout (2x2 for the shifted iterations) and 4 iterations of the above algorithm

Results

Using the error comparison tool, we see that the iterative solution only deviates from the exact one by more than 1.5 px for a handful of tile pairs, all of which contain more or less only resin. Otherwise, the iterative solve looks pretty good to me. In particular, the global coarse block solve has errors well below 1px after the first two runs.

Interestingly, the exact solve seems to have a very visible (and probably wrong) drift in the last few layers, which the iterative solve doesn't.

[trautmane]

Looks good - I made some nice-to-have-but-not-essential comments ...

[trautmane]

Just some minor nits (really :) with one logic change.

[minnerbe]

Thanks for the review! I'll try to incorporate your suggestions later today, but if I don't get around to do it and it's time critical, feel free to edit the code yourself.

[minnerbe]

Thanks for applying the changes yourself, @trautmane ! I was completely swamped the last days...