Fix #1274: fix PadIfNeeded to only pad (no crop)

[abhayrajjais01]

Summary

This PR fixes Issue #1274 where PadIfNeeded was incorrectly cropping images larger than the target size because it was aliased directly to Kornia.PadTo.

Changes

• New PadIfNeeded Class: Implemented a custom PadIfNeeded augmentation (inheriting from GeometricAugmentationBase2D) that:
  • Checks if the input image dimensions are smaller than the target size.
  • Pads only if needed.
  • Returns the identity (unchanged image) if the image is already larger than the target size.
• Added PadTo: Added a new key "PadTo" to _SUPPORTED_TRANSFORMS that maps to the original Kornia.PadTo behavior. This preserves backward compatibility for users who relied on the cropping behavior.
• Tests: Added a new regression test test_pad_if_needed in tests/test_augmentations.py to verify:
  • Small images are padded correctly.
  • Large images are NOT cropped (the fix).
  • PadTo still works as expected (legacy behavior).

Verification

Ran tests via Docker:
docker compose run --rm deepforest pytest tests/test_augmentations.py
Result: All 28 tests passed.

I have taken help of ai to understand the issue.

[Prathamesh8989]

Hey @abhayrajjais01
I just went through your code — great job overall! 🙌

I noticed one small thing that could potentially be improved.

Padding Behavior

On line 135, the current implementation is:

return torch.nn.functional.pad(
    input,
    [0, pad_width, 0, pad_height],  # [left, right, top, bottom]
    mode=flags["pad_mode"],
    value=flags["pad_value"],
)

This applies padding only to the right and bottom of the image.

Why This Matters

Padding on only one side can cause a noticeable shift in the feature center, which may affect how the neural network perceives spatial information.

Using symmetric padding (padding equally on all sides) helps:

• Keep the original image content centered
• Preserve spatial alignment of features
• Reduce unintended bias introduced by directional padding

Suggested Improvement

Consider updating this to use symmetric padding, where padding is distributed evenly across left/right and top/bottom (when possible). This can improve consistency and stability during training and inference.

Let me know what you think — happy to help iterate on this 👍

[abhayrajjais01]

thank you @Prathamesh8989 for the suggestion