Fast Continuous Max-Flow Algorithm to 2D/3D Image Segmentation

Fast continuous max-flow algorithm to 2D/3D image segmentation developed in matlab, C and GPU

Description

This software implements the fast continuous max-flow algorithm to 2D/3D image segmentation. It provides three implementations: matlab, C and GPU (cuda based). All the source files are provided. So it is easy for you to incorporate the algorithms in your applications.

Image segmentation can be modeled as computing the min-cut in a spatially continuous setting. The continuous max-flow formulation is dual/equivalent to such continuous min-cut problem. On the other hand, it also leads to a new fast algorithm in numerics, i.e. the fast continuous max-flow algorithm. More details are given in the document CMF_README.pdf. Wish this software would be helpful for you and your works!

For multi-region image segmentation, the fast continuous max-flow algorithm can be downloaded by:
http://www.mathworks.com/matlabcentral/fileexchange/34224
which computes multiple segments simultaneously!

Email Jing Yuan ([e-mail]) for the questions, comments and suggests. You are welcome to propose your ideas of improvements. Any contribution will be kindly acknowledged in CMF_README.pdf.

• Required Products: MATLAB
• MATLAB release: MATLAB 7.10 (2010a)
• Other requirements: any C compiler, CUDA C compiler

[Jing Yuan]

The constant BLOCK_SIZE determines the total number of threads when the kernel is launched. For the 2D case, it is 16; so 16^2 = 256 threads are launched. For the 3D case, it is 8; so 8^3 = 512 threads are launched.

Please check the GPU you are using and choose a proper value for BLOCK_SIZE.

[Jing Yuan]

The 3D data file *.nii is written in the format of nifti. You can open it by the function load_nii(), which can be downloaded by the matlab toolbox: http://www.mathworks.com/matlabcentral/fileexchange/8797 .

[Jing Yuan]

The time unit using in programs is actually mili-sec., not sec.

For example, if the program shows:

Computing Time for max-flow = 30.0000 sec.

it is actually 30 mili-second, not 30 seconds.

[Martin Rajchl]

 

[Abraam Soliman]

 

[gargoyle]

hi, I have 3d images of cells which i need to process over a time axis.
however, i dont have an Nvidia card (but a ati mobility radeon hd 4670). from what i understand after a brief search online is that i cannot use the CUDA with this processor... is there a work around for this? is there perhaps a way that this code can we written for ati cards? I'm willing to do it although i will need some coaching since i have never worked with GPUs before.

[Jing Yuan]

To gargoyle

For AMD's GPU, you can use OpenCL and revise the codes with OpenCL programming. OpenCL tends to be a GPU programming standard in the future, which both NVidia and AMD support. It has a very similar syntax as CUDA. I'll also rewrite the codes by OpenCL in the future.

You can check this link for an overview of the two GPU programming CUDA and OpenCL:
http://www.streamcomputing.eu/blog/2011-06-22/opencl-vs-cuda-misconceptions

[Eran Ukwatta]