[discussion] [wip] Move/adopt to torchvision frequent utils from detectron2 (and potentially other frameworks)

🚀 The feature

I propose to open the discussion and collect in this issue some discrepancies or duplicate functionalities I found between detectron2 and torchvision.

detecton2 / mmdet / kornia are now quite old and established, so one can discuss of adopting well-designed / widely-used util functions and Modules from these frameworks to torchvision, so that there's less reimplementation and more established idioms.

• move CUDA and C++ ops from detectron2: https://github.com/facebookresearch/detectron2/tree/main/detectron2/layers/csrc: ROIAlignRotated+box_iou_rotated+nms_rotated, deformable convolutions (already moved to torchvision?). Moving them to torchvision would also allow to have a workable python-only version of detectron2
• fast evaluators from detectron2: coco format evaluator / lvis evaluator?
• paste_masks_in_image - already a version exists in torchvision but with slightly different API; and some other mask_ops?
• simple colormap generation
• instance / box visualizer: https://github.com/facebookresearch/detectron2/blob/main/detectron2/utils/visualizer.py
• exif parser / apply exif orientation: https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/data/detection_utils.py#L119
• focal loss from fvcore https://github.com/facebookresearch/fvcore/tree/main/fvcore/nn and other utils from fvcore?

Motivation, pitch

N/A

Alternatives

No response

Additional context

No response

This is a broad discussion/WIP issue with no single target file or test. Start by comparing the listed detectron2 paths, fvcore utilities, and existing torchvision APIs; the work is done only after a specific utility or module is selected, scoped, and its expected behavior and tests are agreed.