2D point radius != None shows up inconsistently in 2D and 3D view

[roym899]

Describe the bug
Logging 2D points to an image plane they show up as huge 3D points (pixels converted to meters?). Things look fine as long as radii=None.

To Reproduce

import numpy as np
import rerun as rr

rr.init("projection", spawn=True)

rr.log_view_coordinates("world/camera", xyz="RDF", timeless=True)

rr.log_pinhole(
    "world/camera/image",
    child_from_parent=np.array([[500, 0, 250], [0, 500, 250], [0, 0, 1]]),
    width=500,
    height=500,
)
rr.log_image("world/camera/image/rgb", np.ones((500, 500, 3)))
rr.log_points("world/camera/image/points", np.random.uniform(0, 500, (30, 2)), radii=1)

Expected behavior

Actual behavior

Desktop (please complete the following information):

• OS: Ubuntu 20.04

Rerun version

rerun_py 0.7.0 [rustc 1.69.0 (84c898d65 2023-04-16), LLVM 15.0.7] x86_64-unknown-linux-gnu prepare-0.7 9cf3033, built 2023-06-16T15:47:32Z

[roym899]

Maybe related to #2279.

[emilk]

It seems all points in 2D views have too small radius, at least on retina displays

[abey79]

This issue has been encountered and reported by some of our users. IMHO, since the 2D point radius correspond to "pixel" units in the 2D space view, and these pixels have a defined size in the 3D space view (on the pinhole widget projection plan), we should be able to adjust the 3D radius to be consistant. That would involve taking into account both the pinhole project itself and the "Image plan distance" parameter.

Expected

Actual

[abey79]

Fun fact: setting the "Image plan distance" to the same value as the pinhole focal yields the correct behaviour:

[SeaOtocinclus]

@abey79 FYI, it seems to happen too if you are logging a Boxes2D, so the problem is most likely happening for more than one archetype.

[abey79]

Indeed, and the glitches are... interesting:

Again, like for points, better handling of "Image plan distance" might improve on this issue:

Code

import numpy as np
import rerun as rr

rr.init("rerun_example_pinhole")
rr.connect()

rng = np.random.default_rng(12345)

image = rng.uniform(10, 30, size=[200, 300, 3])
rr.log("world/image", rr.Pinhole(focal_length=300, width=300, height=200))
rr.log("world/image", rr.Image(image))
rr.log("world/image/point", rr.Points2D([50, 70], radii=5))
rr.log("world/image/box", rr.Boxes2D(centers=[100, 100], half_sizes=[20, 30], radii=2))