[Rtk-users] More errors with 360 projections than 180 for SART

[anais.capouillez at student.uliege.be]

Hi,

Thank you, this is probably the reason.

Anaïs

----- Mail original -----
De: "Simon Rit" <simon.rit at creatis.insa-lyon.fr>
À: "Chao Wu" <wuchao04 at gmail.com>
Cc: "anais capouillez" <anais.capouillez at student.uliege.be>, "rtk-users" <rtk-users at public.kitware.com>
Envoyé: Jeudi 12 Avril 2018 07:50:54
Objet: Re: [Rtk-users] More errors with 360 projections than 180 for SART

Hi Anais,
The suggestion posted by Chao might be helpful to improve both results. My
intuition is that this is caused by the fact that SART has no
regularization and it is know that it will start to diverge from the
solution after a few iterations. Setting the number of iteration is a way
of avoiding divergence.
Now, if you double the number of projections, I would say that one
iteration goes two times faster to the solution. So I would compare a
halved number of iterations when doubling the number of projections. I
would suggest to use regularized conjugate gradient and to compare the
results at convergence instead...
I hope this helps,
Simon

On Wed, Apr 11, 2018 at 9:39 PM, Chao Wu <wuchao04 at gmail.com> wrote:

> Hi Anais,
>
> I am travelling so I cannot review your images, but please check if the
> following applies to you:
> Open issue: https://github.com/SimonRit/RTK/issues/151
> Related thread: https://public.kitware.com/pipermail/rtk-
> users/2018-February/010575.html
>
> Regards,
> Chao
>
> 2018-04-11 15:37 GMT+02:00 <anais.capouillez at student.uliege.be>:
>
>> Hi,
>>
>> I'm reconstructing a homogeneous cylinder. My problem is that I have
>> better results for 180 projections, than for 360 for SART (with default
>> parameters), and the same geometry (except the number of projections).
>>
>> I don't understand why adding projections adds more errors, since it
>> should be better (and for other algorithms this is indeed better with 360
>> projections).
>>
>> With 180 projections, the mean error on the cylinder is 1,6% , and the
>> maximal error is 21,6%. While, with 360 projections, on the cylinder the
>> mean error is 6,4%, and the maximal error is 169%.
>> Even if I don't take into account voxels close to edges, there are more
>> errors with 360 projections.
>>
>> Is there a reason why I obtain worse results with 360 projections?
>>
>> Since my files are too big, I added them here
>> https://drive.google.com/drive/folders/1AYumjqNGUr8DZ0kxIbF7
>> Ab1xet7srk-z?usp=sharing. I also added a file with the commands I used.
>>
>>
>> Thank you very much.
>>
>> Anaïs
>> _______________________________________________
>> Rtk-users mailing list
>> Rtk-users at public.kitware.com
>> https://public.kitware.com/mailman/listinfo/rtk-users
>>
>
>
> _______________________________________________
> Rtk-users mailing list
> Rtk-users at public.kitware.com
> https://public.kitware.com/mailman/listinfo/rtk-users
>
>