What does this PR do?

Previously, box postprocessing worked properly with letterboxed inputs, but mask postprocessing wasn't properly accounting for / cropping out the black bars introduced by letterboxing. Now it does. The logic used is similar to what exists like 20 lines above for the un-letterboxing of the box predictions.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:
This turned out to be the correct fix (or, well, the first part of an even more correct fix) for what was addressed by this roboflow-train PR. My test case was the same — I saw squished masks before the change, and now the masks aren't squished anymore:

evaluated on main branch of inference:

evaluated on this branch of inference:

I also decided to check a dataset where the black bars are on the top and bottom instead of the sides, just to be sure.

evaluated on main branch of inference:

evaluated on this branch of inference:

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

N/A