Object segmentation: defining the object of interest

This question might be a bit more broad.

In case of object segmentation, for instance my object of interest is a person in the image, so before performing the segmentation on the person, how can I identify that person as the object of interest? What I mean to say is, I don't want to segment any other object in the image except the person. How can I do that? I know there's template based approach, however, I was wondering if this can be done using predefined thresholds that defines the sizes of different parts of the body and processes only those parts further?

Answer 1

[Dima Lisin]

Hi Mona,

You can use vision.PeopleDetector from the Computer Vision System Toolbox to give you the rough region of interest. Then you can run a segmentation algorithm on that.

[mona]

Interesting. Is this built-in algorithm more feasible compared to other method. I mean to say, is this algorithm statistically used more than other methods for detection purpose?

Also, I think there will be no need to impose global constraints related to the body size, right?

[Dima Lisin]

vision.PeopleDetector uses Dalal-Triggs HOG-SVM, which is a "classic" people detection algorithm. It is no longer the best, but it is what everyone still compares to.

Answer 2

[Image Analyst]

Well mona, you'd have to segment those features and measure them and throw out the ones you don't want. For example if you want tall men on the beach with small waists and big biceps, you have to first find all the men, then measure their heights, waists, and biceps and then use ismember() to extract out those that meet your criteria. See my image segmentation tutorial for an example: http://www.mathworks.com/matlabcentral/fileexchange/?term=authorid%3A31862

[mona]

This brings down to the question if there are any global constraints in case of human images that are universally used in segmentation algorithms?

[Image Analyst]

I don't know, but if there are they'd be listed here: http://www.visionbib.com/bibliography/contents.html