[Insight-developers] ITK wrapping status

[Michal Suchanek]

2010/10/22 Gaëtan Lehmann <gaetan.lehmann at jouy.inra.fr>:
>
> Le 22 oct. 10 à 19:35, Michal Suchanek a écrit :
>
>> 2010/10/22 Gaëtan Lehmann <gaetan.lehmann at jouy.inra.fr>:
>>>
>>> Le 22 oct. 10 à 17:30, Michal Suchanek a écrit :
>>>
>>>> 2010/10/22 Gaëtan Lehmann <gaetan.lehmann at jouy.inra.fr>:
>>>>>
>>>>>>
>>>>>>
>>>>>> Also the filter intentionally uses UL output to get the best possible
>>>>>> resolution and the image may lose detail when converted to UC.
>>>>>
>>>>> you shouldn't loose anything when converting to UC with the relabel
>>>>> filter
>>>>> if there are less than 256 zones identified.
>>>>> If UC is not enough, there is also US (16bit) available by default.
>>>>>
>>>>>>
>>>>>> The obvious choice of filter for this job it RescaleIntensity but that
>>>>>> is not wrapped with inputs and outputs suitable for this example, nor
>>>>>> is RelabelComponent.
>>>>>
>>>>> RescaleIntensity is NOT well suited for this task, beacause it can
>>>>> merge
>>>>> several labels in a single one. RelabelComponent is the filter you
>>>>> want.
>>>>> RelabelComponent should be wrapped with the necessary types to convert
>>>>> UL
>>>>> pixel type to UC or US.
>>>>
>>>> UC  is not wrapped by default so there are no UC filters at all.
>>>>
>>>> US will likely not work with most image writers either but at least
>>>> the modified JPEG should support it.
>>>
>>> UC is always wrapped for the writer, even when not selected in cmake
>>> options.
>>
>> Yes, it is supported by the writer but not by any filters so you
>> cannot convert it to the writer format.
>>
>
> of course you can do the conversion - it would be a complete waste of space
> and build time to have it for the writer without that!
> CastImageFilter is there for that.

CastImageFilter may convert just about anything to anything but unless
it has changed from last time I read the docs it does not ensure such
conversion gives any data at all. Thus an image which has been
previously normalized would produce a black image when casted to
integer type.

More often than not using CastImage is a complete waste of code lines
and processor cycles.

>
>>> US is well supported by tif and png.
>>
>> It is. I get a ~ black output image.
>>
>> While the RelabelComponent filter may be better for preserving
>> information (zones) in the image regardless of intensity (which is
>> artifical on segmentation output anyway) the RescaleIntensity filter
>> is better suited for a demo because it provides the user with
>> segmentation results visible in a plain image viewer (which does not
>> do the equivalent of RescaleIntensity as default preprocessing).
>
>
> You may want to produce a colored output instead of an almost black image
> then.
> LabelToRGBImageFilter is able to convert your label image in a more human
> friendly, colored image.

Yes, that's good update of the example to also show off new features.

It somewhat hides the mechanics of the watersched filter, though.

Thanks

Michal