Batched box scale

[vadimkantorov]

🚀 The feature

Frequent utility in postprocessing box detections:

def batched_box_scale(boxes, image_shape, target_image_shape):
  fx = target_image_shape[..., -1] / image_shape[...., -1]
  fy = target_image_shape[..., -2] / image_shape[..., -2]
  return boxes * torch.stack([fx, fy, fx, fy], dim = -1).to(boxes.device)

Motivation, pitch

This is a frequently needed utility to bring back computed detections to a common / original coordinate frame.

cc @vfdev-5 @datumbox

[fmassa]

Hi @vadimkantorov

This is planned with the revamp of the transforms. The RFC is in pmeier/torchvision-datasets-rework#1 , and we are starting to integrate it in #4721

Let us know what you think about it

[vadimkantorov]

What do you mean by transforms? For me, this kind of utilities have nothing to do with augmentations. Probably, I misunderstand the term

[fmassa]

I meant this will be handled by

# boxes is a custom object that is defined in #4721
boxes : BBox

# then the transforms will know how to resize them
T.resize(boxes, (50, 20))

[vadimkantorov]

I only hope that there will also be functions accepting bare tensors - this would be more versatile and not require buying into new abstractions for porting existing code bases

[fmassa]

@vadimkantorov new abstractions with transforms are in #4861

I hope they will be useful and would encourage users to use them