Fix presence of segmentation images and tensor causing generation of root space view

[Wumpf]

What

• Fixes Bad heuristic results in adding inappropriate segmentation to 3D scenes #3677

Definitely need to come up with something better to steer the heuristics - we made good strides in space view refactor and recent advances in System heuristics, but there's still a lot to be desired!

In the test scene there's still the issue of a unexplicably large accumulated bounding box which causes the camera to zoom out a lot. This is most likely caused by an initially bad heuristic for the frustum length. I suspect this is a problem we already had in 0.8 and might be fairly hard to solve until we fixup our blueprint & object property story.

Before:

After:

Test code:

"""Log a pinhole and a random image."""
import numpy as np
import rerun as rr

rr.init("rerun_example_pinhole", spawn=True)
rng = np.random.default_rng(12345)

image = rng.uniform(0, 255, size=[3, 3, 3])
rr.log("world/camera/image", rr.Pinhole(focal_length=3, width=3, height=3))
rr.log("world/camera/image", rr.Image(image))

image = rng.uniform(0, 255, size=[3, 3, 1])
rr.log("seg", rr.SegmentationImage(image))

Also made sure this doesn't regress the fix in #3583 (which touched this code last!)

Checklist

• I have read and agree to Contributor Guide and the Code of Conduct
• I've included a screenshot or gif (if applicable)
• I have tested demo.rerun.io (if applicable)

• PR Build Summary
• Docs preview
• Examples preview
• Recent benchmark results
• Wasm size tracking

[emilk]

So is_interesting_space_view_at_root used to return true for segmentation images, but I don't understand how that would put it in the same spaceview as an entity with a different root?

[emilk]

Will this mess things up for non-2D tensors?

[Wumpf]

I think you're right. The old code did this:
https://github.com/rerun-io/rerun/blob/release-0.8/crates/re_viewport/src/space_view_heuristics.rs#L76

[Wumpf]

I'm a bit concerned now. I could do the 2dness check on top and be mostly with the old behavior but this has slight regression potential on the fix from #3583
Well, let's try it out!

[Wumpf]

Actually it would be probably best then to check for Image and SegmentationImage indicators.

[Wumpf]

yes. This seems safest, easiest and passes both repro scripts

[Wumpf]

but I don't understand how that would put it in the same spaceview as an entity with a different root?

Root is the root here, i.e. /. If we decide to create a Space View there we end up deciding it must be 3D and then things go awry

[emilk]

scripts/run_all.py looks good