Replace Albumentations with Kornia for training transforms

[robmarkcole]

Search before asking

• I have searched the RF-DETR issues and found no similar feature requests.

Description

RF-DETR currently depends on albumentations for training resize/augmentation, but Albumentations was archived on July 10, 2025 and is now read-only. This codebase already support Kornia, so we should consider moving all training transforms to Kornia and removing Albumentations.

Benefits:

• Removes dependency on an archived project.
• Avoids pulling OpenCV through Albumentations.
• Unifies CPU/GPU augmentation around one backend.
• Reduces dependency and maintenance surface.

Acceptance criteria:

• Training no longer requires albumentations.
• COCO/YOLO detection and segmentation transforms still work.
• Resize, multi-scale, square resize, masks, boxes, and normalization retain parity.
• Tests cover the replacement paths.

Use case

No response

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!