Increase speed of imwarp

Dear Matlab community

I have been using imwarp to spatially transform CT-images along with an upscaling step. I would like to optimize my code and I've found that imwarp is taking a lot of time for my needs. Since I need to upscale one of the images by about 2x2x3 in the x,y, and z direction, respectively, the function ends up taking around 10 minutes per image. After doing a profile check it is actually interp3 the one function inside imwarp that consumes most of the time. I wonder if there are ways to speed imwarp like using parallel computing or other alternatives.

Best regards, Jaime

[Igor Reidler]

Hi Jaime, I'm facing a similar situation. Was wondering if you had any progress using parallel computing toolbox, GPU computing, or other?

Thanks, Igor.

Answer 1

[Matt J]

IMWARP seems like the wrong tool for upsampling images. griddedInterpolant is probably smarter, in the sense that it knows that the upsampling can be done separably, and hence more efficiently:

      [m,n,p]=size(CTVolume);
      gridVectors={linspace(1,m,2*m),...
                   linspace(1,n,2*n),...
                   linspace(1,p,3*p)};

      F=griddedInterpolant(CTVolume);

      newImage=F(gridVectors);

Answer 2

[Kees de Kapper]

Dear all,

I know this thread is a bit "old" but I've got a similar question. I'm using imwarp for extracting a 2D image from a 3D dataset but with the use of a 3D transformation.

What I experienced is that the speed of imwarp does not depend on the size of the resample output but on the size of the input image. This surprised me a lot and implies a very inefficient implementation.

I found by searching the web, the function tformarray that I might use as well. Using this function the calculation time is reduced with a factor more than ten. I find this very surprising as I expect the imwarp function is build on top of the tformarray-function (correct me if I'm wrong). Do I make a mistake?

The disadvantage is that when using tformarray, I have to construct a tmap_B that reflects the pixel index corresponding to the axis coordinates for both input and resample axis. Although this might be trivial, is there a in-build Matlab function that can do the job?

I'm looking forward to your comments.

Many thanks in advance,

Kees