Centerpoint test discrepancies for 3D downsampling #710
Description
One of the test conditions in tests/test_downsample.py is that the intensity at the center point of the volume should be conserved.
According to our centering conventions this is the point v[L // 2, L // 2, L // 2]
We are not achieving this to within the dtype tolerance for our 3D downsampling code. This is a problem inherent to the original method, which is identical to the one introduced in #689 (this PR added the tests that exposed this problem)
Current progress on investigating: #689 (comment)
Update on tolerance level for each 3D downsample case:
Changing the anisotropy or the max/min sigmas shuffles things a bit but in general we're seeing that even to even is by far more accurate than the other 3 cases.
With current sigmas
Even to Even (64,32): 5e-5
Even to Odd (64,33): 1e-2
Odd to Even (65,32): 5e-2
Odd to Odd (65,33): 5e-2
Anisotropy Removed:
Even to Even (64,32): 5e-5
Even to Odd (64,33): 5e-2
Odd to Even (65,32): 1e-2
Odd to Odd (65,33): 1e-2
Large sigmas removed (0.3)
Even to Even (64,32): 5e-5
Even to Odd (64,33): 5e-2
Odd to Even (65,32): 1e-2
Odd to Odd (65,33): 1e-2
Small sigmas removed (0.1)
Even to Even (64,32): 5e-5
Even to Odd (64,33): 1e-2
Odd to Even (65,32): 1e-2
Odd to Odd (65,33): 1e-2