Re: ENH: add bounding_box function to SciPy.ndimage

Ralf Gommers wrote:
> On Mon, Mar 20, 2023 at 2:06 PM [email protected] wrote:
> > Link to PR: https://github.com/scipy/scipy/pull/18170
> > I would like to propose a new function for the ndimage measurements
> > module: `bounding_box()`
> > This function finds the bounding box of the nonzero elements in the input
> > array. It accepts any dimensionality and dtype input.
> > Why do I think this is needed?
> > It is true that same result can be obtained by:
> > 
> > numpy: `nz=np.nonzero(arr)  ; nz[0].min(), nz[0].max(),
> > 
> > nz[1].min(),....`   : SLOW
> > 
> > scipy: `ndi.find_objects(arr!=0,max_label=1)`   : Faster, implemented in
> > 
> > C
> > I actually implemented `bounding_box` simply by adapting the existing
> > implementation of `find_objects`, so the logic is fundamentally the same.
> > However, it runs significantly faster - namely between 10% and 30% faster
> > than `find_objects`, per my tests locally.
> > This is mainly due to the fact that `bounding_box` runs *in-place* on the
> > input array, while running `find_objects(arr!=0) requires allocation of a
> > new array, to store the boolean result of `arr!=0`, which is then passed as
> > the argument to `find_objects`.
> > Of course, avoiding this allocation also means memory footprint is
> > significantly reduced, and in case of large multidimensional arrays this
> > can be crucial.
> > The only similar functionality I have found in Python libraries is
> > `PIL.Image.getbbox()` :
> > https://pillow.readthedocs.io/en/stable/reference/Image.html#PIL.Image.Image...
> > This is limited to 2D images.
> > Thanks, this seems like a reasonable function to add to me. The one
> question that came to mind is whether the functionality should be supported
> for floating point dtypes, or only for bool and integer dtypes. Your
> implementation uses `!= 0`, and that may not work as expected for -0.0
> values for example. In general these kinds of exact equality checks for
> floats should be avoided.
> Cheers,
> Ralf

Thanks for your feedback, Ralf.

Your question regarding floating point types is an important one.
One could make the claim that comparing `!=0` does make sense in some cases for this function, suggesting that a value that is in any way different from `np.zeros([...])`, is "nonempty".

Maybe we can address the more general case by adding an optional argument, specifying the max absolute value to be considered "zero" in the floating point case, e.g:
`ndi.bounding_box(arr, epsilon=1e-6)`

Tomer
_______________________________________________
SciPy-Dev mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
Member address: [email protected]