2D scene does not take point radii into account when calculating bounding box

[roym899]

Describe the bug
Visualizing 2D points does not seem to work (with Python at least), while docs suggest 2D or 3D points should work.

To Reproduce
Steps to reproduce the behavior:

1. Modify code from docs example

import rerun as rr
import numpy as np

rr.init("my data", spawn=True)

num_points = 100
dim_points = 2

positions = np.zeros((num_points, dim_points))
positions[:,0] = np.linspace(-10,10,num_points)

rr.log_points("my_points", positions=positions, radii=0.5)

2. Run the code

Expected behavior
Should show a scatter plot.

Desktop (please complete the following information):

• OS: Ubuntu 20.04

Additional context
Python 3.8.11, rerun-sdk 0.3.0

[Wumpf]

Thanks for the detailed report with repro!

[jleibs]

I believe this is related to #1113

This is another (much more serious) case of the bounding box being too small for the content.

One of the easiest (though admittedly hacky) ways to force a larger bounding box is to log one explicitly with a rect.

For example, if I use:

rr.log_rect("bounds", [-10, -10, 20, 20])
rr.log_points("my_points", positions=positions, radii=0.5)

Then the points become visible:

[emilk]

Seems like the points are somehow missed by the bounding box calculations, but I don't see how they could be… weird.

[jleibs]

I don’t think they’re missed. They just don’t have their radius taken into account. These points all lie on a horizontal line and so I suspect the bounding box here ends up having zero height.

[emilk]

Ah right, that makes sense. Sounds like an easy enough thing to fix then.

[jleibs]

This turns out to get quite complicated quite quickly -- not actually a good first issue. I started working on it here: #1658

Although that kind of solves one edge case there is a long tail of issues such as supporting auto for points, handling lines, appropriately, and the fact that the future staging belt design means accessing this data from the builder may be an anti-pattern we should be moving away from.

The better short-term solution here might be to give manual control of changing the bounding box for 2d views as a scene property so we have an escape hatch.

[abey79]

The map view suffers from the same issue for the same reason. It's probably even more complicated to solve, as the scene unit to ui point relationship varies based on the latitude (in addition to whatever current "camera" setting the view has).