Add custom Augmentations - Albumentation Wrapper

[panagiotamoraiti]

Description

This PR adds custom Albumentations wrappers and additional augmentation classes to the rf-detr repository, addressing issue #81. The changes enable seamless integration of Albumentations transforms with bounding box support, as well as custom resizing and cropping augmentations missing from torchvision. A RandomSelect wrapper is also introduced to probabilistically select between two augmentation pipelines, increasing augmentation diversity.

This enhancement improves flexibility and configurability of data augmentations for object detection training pipelines.

Dependencies:

• Albumentations library

Type of change

Please delete options that are not relevant.

• [Augmentations]New feature (non-breaking change which adds functionality)

How has this change been tested, please provide a testcase or example of how you tested the change?

Verified integration by running training with new augmentations enabled.
Additionally, original vs. augmented images were saved along with the bboxes for further inspection.

You can check an example below:

The augmentations are loaded from a config file.

[Akhp888]

Hello @panagiotamoraiti

Thanks for working on this PR , I agree this is a very important part missing in the current setup .
Said that based on your changes i would suggest to also add copy-paste and mosaic as well to the augmentation options.

Thank you

[panagiotamoraiti]

@Akhp888 Nice suggestions! I can definitely add these additional augmentations.

[raphael0202]

Any news on this? It would be really valuable to have albumentation integration to RF-DETR!

[panagiotamoraiti]

@raphael0202 I think someone responsible will evaluate the pull request. Till then you may use the code from this branch.

@isaacrob-roboflow @SkalskiP Any news about this issue?

[mkrupczak3]

👀