Comments on: Spatial transformations

By: Steve

Steve — Mon, 06 Jul 2009 15:26:14 +0000

Shai—I don’t think you’ll be able to establish a direct connection between the GPU and a MATLAB figure window. And from what I’ve heard, the bus bandwidth is indeed a problematic bottleneck in GPU-based computation. That is, if you want to get the result back into MATLAB. ;-)

By: Shai

Shai — Sun, 05 Jul 2009 11:38:02 +0000

I am also very concerned about the performance of tformarray.
Is there a way to harness GPU capabilities to assist in this case? GPU is very much tuned to performing all sorts of transformations on images (at least) can it be used to perform image/array transformations in Matlab as well?
I was thinking of creating a figure with opengl renderer, define an image as a graphical object and apply the transformations to this object. The bottle neck here (as far as I can tell) is reading from the GPU to main memory for further processing. Can you comment about this?

By: Steve

Steve — Thu, 15 Feb 2007 22:15:18 +0000

S.Vibin - Take a look at this contribution on the MATLAB Central File Exchange.

By: vibin.S

vibin.S — Tue, 13 Feb 2007 05:16:47 +0000

Respected Sir
I am doing X-ray CT reconstruction part,I having some problem in 3D reconstruction, i,e how to stacking the data using MATLAB code. If You have some idea kindly mail me..
Thank you
by
S.Vibin

By: Steve

Steve — Wed, 22 Nov 2006 13:10:34 +0000

Abe – I think you could probably set up a MATLAB timer object to update your waitbar.

By: Abe

Abe — Wed, 22 Nov 2006 07:10:50 +0000

Steve, thanks for the time to confirm that I am getting the expected behaviour. The memory usage is definitely acceptable, no swapping. It’s just the time it takes which is a bit long. I was using other softwares to do the same 3-D rotation and like what Philip has got, a few times faster. Anyway, I guess it is probably not fair to compare tformarray to other softwares out in the market which is binary-coded.

By the way, is there any way I could continuously update my waitbar to reflect at least some “progress” going. It looks as if the GUI freezed for a whole minute or two before tformarray returns, which isn’t very good in terms of usability/presentability. :)

By: Steve

Steve — Tue, 21 Nov 2006 01:12:36 +0000

Abe – I profiled a similar case on my laptop, and it ran in about 100 seconds. 75% of that time was in the MEX code for performing separable resampling, and there was no apparent swapping going on. I don’t see anything obvious to suggest to you for speeding up this operation.

By: Abe

Abe — Fri, 17 Nov 2006 04:11:34 +0000

I’ve been using tformarray to rotate a entire stack of 3-D CT images (dimensions~512x512x250).
rot_bone_mask = tformarray(bone_mask, tform_trans, R, [2 1 3],[2 1 3], floor([max_x max_y max_z]), [], []);

I’ve segmented it and converted it to uint8, hopefully reduce the memory requirements of it. However, it seems that the tformarray takes quite a fair bit of time to run, around 1-2 minutes. I am really interested in hearing your opinion about the performance of tformarray, or how to make the best use of it, as the performance I get here is far for satisfactory. Thanks for your input.

By: Steve

Steve — Fri, 27 Jan 2006 14:20:21 +0000

Philip – UData, VData, XData, and YData parameters specify the locations of pixel centers, not pixel edges. I’ll get into performance issues later.

By: philip batchelor

philip batchelor — Fri, 27 Jan 2006 12:05:04 +0000

I’m certainly interested in this topic, it took me a while to understand why translating an image doesn’t appear to to do anything ;) and some colleagues have also come to me quite puzzled…I’m still not sure what’s the most correct: I set ‘xdata’, ‘ydata’ using the dimensions of the input image, as I want the transformed one in the same field of view in general, but should it be …,’xdata’,[1,n1], or or [0.5,n1-0.5] ?
(seemed more correct when I tried to overlay lines to show a defomration).

I’d be also interested in the issue of speed for tformarray, I have been running code optimizing over 3D spatial transforms, and the profiler shows most of the time is spent there. For the moment I use an external library (mex+vtk) to perform these spatial transforms, with a speedup of ~10. It seems to me the time lost is in interpolation-reslicing parts of the code, it would be great to be able to use tformarray though, as my code doesn’t have the flexibility of ‘custom’ transforms, I need to write new c++/mex code for every new type of transform. So, well, in summary, any design issue related to tformarray I’d be curious to hear about…