bounding_box = (
    bounding_box.clone()
    if format == features.BoundingBoxFormat.XYXY
    else convert_format_bounding_box(bounding_box, old_format=format, new_format=features.BoundingBoxFormat.XYXY)
)

IMO, this is too bulky. I do not mind to have copy in convert_format_bounding_box and reduce lines of user's code...

@vfdev-5

I do not mind to have copy in convert_format_bounding_box and reduce lines of user's code...

Same here with the focus on consistency. If one convert_* kernel or dispatcher has a copy flag, every should have it. Currently we have copy: bool = True, which has a runtime implication unless someone explicitly sets the copy=False in case no copy is needed. Meaning convert_dtype of #6783 (comment) will be slower by default than before for mostly no reason. This PR basically hard-codes copy=False.

Vasilis argument was that besides for convert_format_bounding_box we aren't using the copy flag anywhere. Even that is only used a handful of times throughout our API. Thus, if we want consistency, it is fairly simple to remove and avoids dead code. I don't want to misrepresent the argument, so I let him chime in here. cc @datumbox

I'm ok with or without the copy flag if we are consistent with it. I'll let the two of you hash it out.

@datumbox @pmeier can we do the following:

• make all conversion ops like _xywh_to_xyxy inplace (without cloning internally) and
• we can apply .clone() for the output of convert_format_bounding_box

bounding_box = convert_format_bounding_box(
    bounding_box.clone(), old_format=format, new_format=features.BoundingBoxFormat.XYXY
)

?

@vfdev-5 that would make our only operator that does in-place ops (outside of the optimizations that have no effect). If we were to go down that route, purely for speed reasons, I would recommend adopting the inplace argument for this one kernel (because it's used so often in conversions). The default should be inplace=False in that case. I would do this outside of this PR though after measuring its speed.