rhoana / fijibento

Tools for aligning very large ssEM image stacks on a compute cluster.

FijiBento

Tools for performing image alignment (mainly for serial section EM) on very large datasets, in parallel on a compute cluster.

Based on Stephan Saalfeld's alignment tools which form part of Fiji and TrakEM2 (least squares registration, and elastic registration etc): https://github.com/axtimwalde/mpicbg And renderer project https://github.com/saalfeldlab/render

fijibento's Issues

Automatic detection of blurred images

Need to automatically detect blurred images and give feedback to the guys that scan as soon as possible.

MFOV identity should be explicit in tilespec

Currently, we have to use either directory or filename information to separate MFOVs. It should be explicit in the tilespec.

Matches file should be indexed and in a fast-parsable format

Hierarchical 2D optimization

First optimize each mfov separately, then "fix" all tiles in each mfov, and then do a global optimization and find a per-mfov transformation (keeping the previous "fixed" tiles transformations).

To reduce the floating point error, each mfov center (with all tiles) should be first transformed to (0,0),
and the result should store both the tile local tile transformation (after the optimization, just the rigid transformation in the tile coordinates), T_tile, and the global translation of each tile to the correct global location, F_tile.
The tile transformation will be based on two transformations: T_tile + F_tile.

The global optimization needs to find a rigid transformation between all mfovs.
To this end, a per-mfov rigid transformation, T_mfov, will be searched for, s.t., T_mfov*T_tile + F_tile will be the tile's final transformation.

Modify the matches viewer to view all matches from all mfovs

Mesh structure should be defined in mesh JSON, not implicit

Test thin plate splines instead of moving least squares

Need to test it on a restricted control point pairs

Automatic detection of tiles or mfovs that have no actual data in them

and ignore these tiles/mfovs

Need tools to gather post-process intermediate files and show statistics

Possible statistics:

• SIFT features count: per mFoV and per Section
• 2D matches: intra and inter mFoVs
• 3D pre-match: #matched mFoVs / #total mFoVs, for each matched mFoV, how far is it's transformation compared to the neighboring mFoVs.
• 3D block-match: #matched points / #total points - per mFoV and per Section.