Bugfixes: Rotation and Scaling

What's Changed

• Add OCR Example by @JosephCatrambone in #4

• Make rotation estimates more conservative for text-in-rectangle. After fix, about 0.25% (1 in 400) text composites are out of bounds slightly on the left of an image. The origin of the issue isn't clear, but I've been fighting this issue for about a month and I'm low on ideas for where the issue appears.

Full Changelog: v0.1.0...v0.1.1

New Truncation Behaviors

Experimentation has shown the need for more granular behavior when text is too long to fit on an image. Before this release, a user could only choose between having an exception raised and returning an empty string. This release adds two new behaviors, long_text_behavior="truncate_then_shrink" and long_text_behavior="shrink_then_truncate", in addition to the existing long_text_behavior="exception" and long_text_behavior="empty". Like the name suggests, truncate_then_shrink will attempt to reduce the length of a string by dropping characters until it fits in an image. This can be useful if you're using the default font-size-choice-behavior (random). The other option, shrink_then_truncate, has proven to be slower for sentences that don't have line-breaks.

Additionally, the new default behavior of text overflow is exception, since silent failures were causing difficult-to-debug downstream errors.

Prerelease: New Return Type (Again)

What's Changed

• pyproject.toml version and GitHub Action versions were out of sync. Resynchronize and re-release 0.0.9.
• Dataset now generates TextOverlayExample in addition to the composite image and text. TextOverlayExample contains 'advanced' details like the name of the font in the image and properties of the distortions.

Full Changelog: https://github.com/JosephCatrambone/PyTorchTextOverlayDataset/commits/v0.0.9