ImageDraw.polygon regression if polygon is a line

[alessiamarcolini]

What did you do?

The behavior of ImageDraw.polygon changed between versions 8.0.4 and 9.0.0 if the polygon is a line.
Could it be related to #5694?

import PIL
import PIL.ImageDraw
import numpy as np

line_vertices = [(1, 0), (1, 0), (2, 0), (2, 0)]
img = PIL.Image.new("L", (6, 5), 0)
PIL.ImageDraw.Draw(img).polygon(line_vertices, outline=1, fill=1)
result = np.array(img).astype(bool)

What did you expect to happen?

This was the result with Pillow 8.0.4

array([[False, False, False, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False]])

What actually happened?

This is the result with Pillow >= 9.0.0

array([[False,  True,  True, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False],
       [False, False, False, False, False, False]])

What are your OS, Python and Pillow versions?

• OS: macOS 12.13.1
• Python: 3.9.7
• Pillow: 8.0.4 and >= 9.0.0

[radarhere]

Testing, the change is because of #5835.

Because that PR was undoing part of #4333, I checked, and yes - the Pillow >= 9 behaviour is also how Pillow behaved <= 7. So I don't think this can be considered a clear-cut regression.

Seeing that aside for a moment, could you give an explanation why you think the current behaviour is not correct? Yes, you're drawing a polygon of zero height, but you're also outlining it, so it doesn't seem unreasonable to me that something is drawn.

[alessiamarcolini]

Thank you @radarhere for the quick reply. I was reporting it not necessarily saying that the new behavior is wrong, just I was not sure which one was the "correct" one. Also, I was not sure if you were aware of this change :)

I spotted this because this example was an edge case for a test in histolab, where we were mostly using it to document what would happen if the polygon was a line. So I'm fine changing the expectation on our end, I just wanted to make sure this was the intended result

[radarhere]

#4333 wasn't intended to change which pixels were filled in, so yes, I'd say that the <=7 behaviour of your code is correct. Meaning that the behaviour >= 9 is also correct.

[radarhere]

So, #4333 unintentionally changed this behaviour. #5835 switched your code back to the original approach, but left the new approach in place when translucency is involved.

I've created PR #6278. It doesn't change the output of your code, but it does fix 1px high polygons when translucency is involved - meaning that even if #5835 hadn't switched back to Pillow's behaviour <= 7, it would now.

[alessiamarcolini]

Thank you @radarhere for the investigation and for confirming that this is the correct behavior!
I will change the expectation of the test on my end 💯