Again fix translate of overlays (this time with tests)

[Czaki]

References and relevant issues

closes #6897

fix again bug from #6633

Description

This PR is fixing again problems with transforms of overlays, this time it contains basic tests.

[Czaki]

@psobolewskiPhD @brisvag During fixing this I hit multiple walls (before I decided to use corner pixels). Could you please try to play with yours datasets and validate if it looks correctly?

@napari/core-devs testing this is not trivial, but looks really required. Maybe one of you will have idea how to improve tests.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.76%. Comparing base (6d207bf) to head (0fafb5c).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6927      +/-   ##
==========================================
- Coverage   92.82%   92.76%   -0.07%     
==========================================
  Files         613      613              
  Lines       55640    55681      +41     
==========================================
+ Hits        51647    51651       +4     
- Misses       3993     4030      +37     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jni]

Sorry, I missed the earlier discussions — can you tell me what a "child node" is in the context of multiscales? Are they tiles? Or...?

[jni]

Adding docstrings to the tests would be really handy. "Issue #6897 showed that we were offsetting overlays for multiscale layers incorrectly because [...]. This test checks against that by creating a multiscale layer then [...], which previously would have done [...] but now does [...]."

[andy-sweet]

@napari/core-devs testing this is not trivial, but looks really required. Maybe one of you will have idea how to improve tests.

Not sure if it's helpful, but in general, I like the vispy integration tests like the ones you wrote. They exercise a critical large integration of napari's code, without bringing in the whole viewer. So I think a little complexity/maintenance is worth it. Will try to take a more detailed look later today or tomorrow!

[Czaki]

Sorry, I missed the earlier discussions — can you tell me what a "child node" is in the context of multiscales? Are they tiles? Or...?

Child nodes are overlays at this moment. I'm not familiar with async rendering.

[brisvag]

Possible followup: I think the "nice" way to solve this would be to have multiscale layers be an custom empty node in the vispy scenegraph with overloaded bounds, so that from the outside it looks like the full normal-sized image. Then, the tiles should be children of this.

[Czaki]

Possible followup: I think the "nice" way to solve this would be to have multiscale layers be an custom empty node in the vispy scenegraph with overloaded bounds, so that from the outside it looks like the full normal-sized image. Then, the tiles should be children of this.

Maybe even have ImageMultiscale and LabelsMultiscale multiscale classes that will be selected in Image.__new__ and Labels.__new__ if multiscale data are passed so we could have a clean independent implementation without multiple if. And then we could implement Vispy visual (or another backend) as you described.

[brisvag]

Sorry, I missed the earlier discussions — can you tell me what a "child node" is in the context of multiscales? Are they tiles? Or...?

@jni, child_node here always refers to vispy nodes that are children of the actual "data" node, i.e: overlays. In the case of multiscale, the data node is the tile. The child is still an overlay. The reason why this needs special treatment is that overlays like bounding box and transform box are relying on the parent node's transforms and bounds() method to know "how big" the data is, and in the case of multiscale this needs to basically undo the tiling step, IIRC.

[jni]

Ok I've updated the code comment to be (I think) a bit clearer, otherwise looks good to me, thank you @Czaki and @brisvag for clarifying for me, and sorry that this fell off my radar.