Skip to content

new multi-SEM layer offset process#218

Merged
trautmane merged 25 commits into
newsolverfrom
newsolver_find_layer_offsets
Jun 13, 2025
Merged

new multi-SEM layer offset process#218
trautmane merged 25 commits into
newsolverfrom
newsolver_find_layer_offsets

Conversation

@trautmane

Copy link
Copy Markdown
Collaborator

To help us discuss changes, I plan to post multiple commits to this pull request as the new process is developed. Review comments are welcome at any time, but I'll add a comment indicating when I think I'm actually done in case you want to wait until then to start reviewing things.

trautmane added 24 commits May 19, 2025 07:39
…use the single best connected tile in each MFOV as a starting location
…sNearestToBox. Previously tiles with more overlap were not necessary considered to be nearer.
…0-ish pixel) differences between x offsets of different MFOVs in the z-1-to-2 test group. Original expectation was that differences between MFOVs in the same layer would almost always be small.
…s. Added --mopMaxAbsoluteMFOVTranslationDelta 150 to test case after learning from Thomas that the distance between 2 MFOVs can vary by up to 12 um = 150 pixels.
…rials work (e.g. between z 9 and 10).

Identify largest MFOV translation delta x and y values and log them.
Increase test case z range to be 1 to 30.
… so that all matches are included in ransac filter, validate offsetStackId at beginning of run
@trautmane

trautmane commented Jun 13, 2025

Copy link
Copy Markdown
Collaborator Author

I plan to merge this pull request into the newsolver branch in a few minutes so that it can be deployed to Janelia's render web services boxes.

Once this is deployed, here is what remains to be done:

  • use RenderTilesClient to render tiles with dynamic downsampled MFOV URLs and save a "hack stack" of the rendered tiles (compute cost: 300 seconds * 76 slots = 7 slot hours * $0.05 = $0.35)
  • generate matches for the dynamic downsampled MFOV stack (compute cost: 2422 seconds * 332 slots = 224 slot hours * $0.05 = $11.20)
  • generate matches for the MFOV "hack stack" (compute cost: 1668 seconds * 332 slots = 154 slot hours * $0.05 = $7.70)
  • compare match results and compute costs:
    • total compute cost with "hack stack" is cheaper: $8.05 vs. $11.20 when using dynamic data
    • is the difference enough to justify adding code to persist hack stacks to Google bucket storage? - yes (also suspect there will be other benefits to having the rendered MFOV s available)
  • patch within-layer connections for any unconnected MFOV tiles using stage coordinates
  • align MFOV stack with translation model
  • apply MFOV tile solve result to each SFOV in the original stack

Most of the TODO items can be done with existing code. If anything needs to be added or fixed, I'll open another pull request for that.

Feel free to add review comments to this pull request - I'll simply address them in a subsequent pull request.

@trautmane
trautmane merged commit 07ce51b into newsolver Jun 13, 2025
2 checks passed
@trautmane
trautmane deleted the newsolver_find_layer_offsets branch June 18, 2025 16:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant