There are many researches which can help you find the best algorithm I suggest the following to to read:

Jason J. Corso , Maneesh Dewan , Gregory D. Hager, Image Segmentation Through Energy Minimization Based Subspace Fusion (2004), N PROCEEDINGS OF INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR)

Qiuxu Li  and jieyu Zhao, MRF Energy Minimization for Unsupervised Image Segmentation,2009 Fifth International Conference on Natural Computation (ICNC 2009), china, http://doi.ieeecomputersociety.org/10.1109/ICNC.2009.354

Good luck

The answer depends very largely on what requirements exactly you have on your segmentation:

* binary segmentation  (one foreground object against some background)

* superpixel segmentation

' 'semantic ' (=object-oriented) segmentation, using priors of how objects look like.

There are many more other variants of problems which can be summarized under the term "segmentation".

The two essential building blocks of a typical (modern) segmentation scheme are

1) an energy function (which defines which segmentation result is 'good' due to agreement between the segmentation and the given image data, in most cases also incorparating a 'penalty' or 'smoothness' term which expresses a priori expectations of how a good segmentation result should look like.

2) an optimization scheme. Currently, graph-cut based methods usually give very good results, and they can guarantee sometimes even (approximate) optimality of the result, in some sense.

For the energy function, I strongly suggest to use a complete statistical model, both for the region contents (the 'texture') as well as for the partition of the image grid into regions (usually some kind of a Markov random field).

For the optimization part, members of the family of graph cut algorithms are often a very good choice.

Recent recommendable work comes from Arbelaez et al.:

for instance "Contour detection and hierarchical image segmentation"

Pablo Arbelaez, Michael Maire, Charless Fowlkes, Jitendra Malik

2011/5, Pattern Analysis and Machine Intelligence, IEEE Transactions on

If you are interested in speed, or in particular if you already have an initial idea of the segmentation result in advance, for instance in sequence segmentation, you could also use a (suitably designed) local greedy optimizer. For instance, you could try out the contour relaxation approach

C. Conrad, M. Mertz, and R. Mester. Contour-relaxed superpixels. In A. Heyden, F. Kahl, C. Olsson, M. Oskarsson, and X.-C. Tai, editors, Energy Minimization Methods in Computer Vision and Pattern Recognition, volume 8081 of Lecture Notes in Computer Science, pages 280–293. Springer Berlin Heidelberg, 2013.

(apologies for recommending work from my group)

Alternatives to graph cut:

Cost volume filtering:

https://www.ims.tuwien.ac.at/publications/tuw-210567

Fast Mean Field:

http://graphics.stanford.edu/projects/densecrf/

Fast Primal Dual:

http://www.csd.uoc.gr/~komod/FastPD/

There are also numerous ways to accelerate graph cut.

Well, most likely you already know, it is the first hit in Google for graph cut c++, but did you try:

http://vision.csd.uwo.ca/code -> Max-flow/min-cut

I use their algorithm (but older open-source version) and it was pretty quick for me.

hi all, 

I am looking at Object Segmentation and using GCMex: http://vision.ucla.edu/~brian/gcmex.html

Maybe I will try one cut by http://vision.csd.uwo.ca/code, and also contour relaxed later as suggested by Rudolf Mester.

There is one idea that I am thinking is to use superpixel in graphcut, anyone has experience in this? I am not aware of anywork on this.

Basically, my aim is to do this real time if possible, while currently it is around 3 sec an image.

Thanks

you can use belief propagation algorithm, ICA, graph Cut, MCMC, SA et al

Hi,

for a overview of existing optimization methods our benchmark paper might help You:

Jörg H. Kappes, Bjoern Andres, Fred A. Hamprecht, Christoph Schnörr, Sebastian Nowozin, Dhruv Batra, Sungwoong Kim, Bernhard X. Kausler, Jan Lellmann, Nikos Komodakis, Carsten Rother : "A Comparative Study of Modern Inference Techniques for Discrete Energy Minimization Problems", CVPR, 2013

Implementations for all methods are available within OpenGM

The choice of the right methods depends on our objective function/model. If You would like to solve this fast You should restrict it to classes which a efficiently solvable or for which efficient and good approximative methods exist.

If runtime is crucial You should also think about efficient building and updating models and using warm start for your solver. Furthermore using Matlab(-interfaces) might be not the most efficient way of implementation even if You use mex-files.