Context

This PR disables integer dtype for rotated bounding boxes. Indeed rotated bounding boxes typically have floating coordinates for vertices. Functions to generate boxes or to transform boxes will typically involves trigonometric sinus and cosinus functions, which does not guarantees the coordinates will be integer. Rounding to the nearest integer can lead to degenerated boxes. Chaining transformation can further exacerbate the degeneration as rounding error will compound. For these reason, we choose to raise an error in the BoundingBoxes constructor if a integer dtype tensor is passed together with a rotated bounding box format. This check is done in the _check_format static function.

Testing

We adapt the tests with the following condition to verify that tests combining integer dtype and rotated bounding box formats are indeed failing.

if not dtype.is_floating_point and (
            tv_tensors.is_rotated_bounding_format(old_format) or tv_tensors.is_rotated_bounding_format(new_format)
        ):
            pytest.xfail("Rotated bounding boxes should be floating point tensors")

We finally remove artifacts in the testing and transforms added to make sure the transforms were compatible with integer dtype for rotated boxes. This includes all the rounding operations as well as epsilon offsets in the testing.

Please run tests with:

pytest test/test_transforms_v2.py -k box -v
...
3230 passed, 5686 deselected, 678 xfailed in 129.91s (0:02:09)