constantinpape / affogato Goto Github PK

The grid graph currently introduces edges for all pair of seeded nodes / pixels.
This introducs O(N_seeded_pixel^2) edges.
We might get away with O(N_seeds^2) edges if we only introduce edges between representatives.

This should def. be viable for repulsive edges, not sure about attractive edges.

From the meeting today: we want to enable interactive mws out of core. High level API discussion:

• segment: just segment everything given the current constraints (respecting locked segments)
  • in core, already there, corresponding to call function of the interactive mws class in prototype
  • out-of-core, prototype exists for a block-wise mws, but cannot ingests user-defined constraints
• segment_current_fov: segment everything in current field of view, apply some "padding" (e.g. by extending two next neighbor in the rag) for fast iteration (both in core / out of core), not implemented yet

Thinking about this more, a challenge for running the interactive mws out-of-core is that (in the current iteration) we need a grid graph built up for the entire image. This happens in memory and is also not economical to do for a very large volume out of core, because we would need to serialize every node (=pixel) and edge (= all local and long-range-offsets).

A potential solution is to never construct the big grid-graph, but instead construct small graphs for the field of view on the fly when a user interacts with the plugin and then use this grid graph to export to the constraint list.
To make this work fully out-of-core, the block-wise mws would then need to be implemented s.t. it computes the grid-graph for its current block and ingests the extra constraints.

I think that this does not need many changes compared to our current implementation, we just need to pass the shape and strides for the full array when computing the grid graph, s.t. the node index is treated globally. For the block-wise mws we need to think about how to handle affinities in-between block boundaries (probably just by adding a corresponding halo).

https://github.com/constantinpape/affogato/blob/master/.github/workflows/build.yaml#L17-L18

There has to be a relabeling after masking, see the semantic branch. If I have time tomorrow morning I will fix it. Otherwise, this will be a reminder ;)

Hi Constantin,

I tried to install affogato on a M1 mac, but there is no native osx-arm64 target. I tried to install it using the x64 emulation and works, but then I have other problems with other dependencies like napari.

Would it be possible (and straight forward) to compile affogato for osx-arm64?

If you need help testing, I have access to an M1 mac.

Best,
Lorenzo

I think compute_weights_and_nh_from_affs does not do what the name suggest.
This seems to be a function tailored for the causal mws.
I suggest we rename it to compute__weights_and_causal_nh_from_affs

Tests fail or even segfault:
https://github.com/constantinpape/affogato/blob/master/src/python/test/segmentation/test_connected_components.py