Inconsistent representation of box in torchvision.ops

Feature suggestion

When checking the how the new operations in the torchvision 0.11 work, I found an inconsistency between different function.

import torchvision.ops as tops
import torch
import torch.nn as nn

a = torch.zeros(1, 1, 8, 8)
a[..., :4, 4:] = 1
bbox = tops.masks_to_boxes((a==1)[0].float()) 
area = tops.box_area(bbox) # the returned area is 9 which should be 16

new_box = tops.box_convert(bbox, 'xyxy', 'xywh') # which returns [4, 4, 3, 3] => expected output [4, 4, 4, 4]

Suggest a potential alternative/fix

This is not actually an error but could cause confusion.

I suggest to make the box representations between different function more consistent.

Start by reproducing the supplied example and inspect the torchvision.ops entry points masks_to_boxes, box_area, and box_convert. Compare their box-coordinate conventions and related documentation or tests; done means the representation is consistent across these operations and the example has a regression test.