The bane of Cell-ACDC: `obj.centroid`

[Teranis]

Description:

So I did a lot of benchmarking, and BY FAR the worst offender when it comes to compute time is obj.centroid when switching slices, and I can only assume frames too! So I had the idea to use cached centroids, where we init all centroids based on acdc_df and then only update when calling update_rp (and if we want to be super efficient, we can also give it a kwarg to only update specific IDs. An implementation solely based on acdc_df completely eradicated any lag whatsoever, which was very promising!

[Teranis]

Additionally, it is accessed in the setannotations part of code. I am unsure why this is even iterated, when I dont have and checked.

[Teranis]

Another idea is to create a worker which takes care of getting rp and precaching centroids, as annotations are more UI "flare" and don't have to be super precisely put.

[ElpadoCan]

Hi Timon, nice investigative work! Amazing how random things are all of a sudden slow when going from 2D to 3D. Another super slow region properties in 3D is solidity. Anyway, pre-caching centroid makes perfect sense then!

However, we need to be careful with detecting that the object changed (hence centroid changed) and new objects. I guess we need an updateCentroids function on every possible edit? To go through every possible function that edits the masks you can search for "Stuff" because that's the keyword I use in the comment before calling the function that stores the edit to enable undo.

Please, do not use workers in separate threads. While they sound cool they are very hard to make completely indipendent and there is always a slowdown of the GUI (see autosave worker) even if it's in theory running in parallel.

[ElpadoCan]

Oh, to Cache centroids we can also subclass skimage RegionProperties class where we reimplement centroid property, add the updatekwarg to the centroid with default to false and when centroid is called it computes only if not cached already or update kwarg is True. Then we need to be careful with where to use update=True

[Teranis]

So for many objects and contour annotations, almost all lag is coming from recreating rp in updateContoursImage. I am wondering if we can either avoid recreating it or at least if you are zoomed in reduce compute time by just getting rp from a small cutout