Suggestions for micro-sam tool

[anwai98]

I received some feedback and requests:

• Possibility to store and export annotation meta-data / macro recordings (regarding the steps for data annotation using automatic / interactive instance segmentation) to ensure reproducibility, and probably mentioning the workflows in publications involving data annotation (wherever relevant)
• Explanation of "embedding file" - some explanation towards what's intended to store here / pass here would help to understand this better.
  • One more possible addition could be, the user isn't prompted correctly on what to do while they use tiling in micro_sam.annotator and not pass the embedding file (the compute image embeddings in .zarr format), it would be nice to lead the users towards how to precompute embeddings (using a link or redirecting to a brief mention of scripts).

Some observations from my side:

• micro_sam.precompute_embeddings - expects -o attribute to store the image embeddings. It would be nice (and probably consistent) to expose the attribute as -e (the same expectation as for annotator_2d, etc., or vice-versa)
• (Minor) micro_sam.annotator_2d - the choice of the model is limited to --model_type. It would be nice to have -m attribute exposed as well.

[anwai98]

Updates based on discussions today:

• Save the information for interactive annotations (if possible, expose a one-click option to export them)
• Advanced addition: To store the projected prompts (preferably in an additional layer, not exposed by default - ideally, users would need to specify some option to access this) to aid interactive improvement in volumetric segmentation.

[anwai98]

Providing a demo notebook for:

• AMG / AIS (automatic segmentation methods)
• Interactive segmentation (using box and/or points)

Hint: something like this - https://github.com/facebookresearch/segment-anything/blob/main/notebooks/automatic_mask_generator_example.ipynb

[constantinpape]

I will close this since most points here are either solved or covered by other issues.

• Possibility to store and export annotation meta-data / macro recordings (regarding the steps for data annotation using automatic / interactive instance segmentation) to ensure reproducibility, and probably mentioning the workflows in publications involving data annotation (wherever relevant)

This is being implemented as part of #408.

Explanation of "embedding file" - some explanation towards what's intended to store here / pass here would help to understand this better.

* One more possible addition could be, the user isn't prompted correctly on what to do while they use tiling in `micro_sam.annotator` and not pass the embedding file (the compute image embeddings in .zarr format), it would be nice to lead the users towards how to precompute embeddings (using a link or redirecting to a brief mention of scripts).

This is a bit outdated. It's now possible to compute tiled embeddings without using the embedding file. In any case, we should improve the documentation on this, but this is covered by other issues already.

• micro_sam.precompute_embeddings - expects -o attribute to store the image embeddings. It would be nice (and probably consistent) to expose the attribute as -e (the same expectation as for annotator_2d, etc., or vice-versa)

  • (Minor) micro_sam.annotator_2d - the choice of the model is limited to --model_type. It would be nice to have -m attribute exposed as well.

Both changes are now implemented. (Together with a better documentation of the parameters of the CLI scripts)

• Save the information for interactive annotations (if possible, expose a one-click option to export them)

  • Advanced addition: To store the projected prompts (preferably in an additional layer, not exposed by default - ideally, users would need to specify some option to access this) to aid interactive improvement in volumetric segmentation.

Also covered by #408.