Machine learning - svm feature fusion techique

Question

for my final thesis, I am trying to create a face recognition system by combining information about color and depth. The first step I took was to rebuild the data head onto a given head model using an iterative nearest-point algorithm. for the discovery phase, I was thinking about using libsvm. but I don’t understand how to combine depth and color information with one vector tag? they depend on the information (each dot consists of color (RGB), information about the depth and quality of the scan). What do you propose to do? something like weighing?

edit: last night I read an article about SURF / SIFT functions, I would like to use them! can it work? the concept will be this: extracting these functions from a color image and depth image (range image) using each function as one vector function for svm?

Answer 1

Concatenation is truly an opportunity. However, since you are working on face recognition with three faces, you should have some kind of strategy regarding how you do it. Face rotation and translation will be difficult to recognize with a “simple” approach.

You must decide if you are trying to perform face detection in general or auxiliary objects. You can try to detect rotation by finding some basic functions (eyes, nose, etc.).

Also, remember that SVMs are essentially binary (i.e., shared between two classes). Depending on your specific application, you will most likely have to use a strategy of several classes ("One against all" or "One against many").

I would recommend doing some literature research to find out how others attacked the problem (a Google search would be a good start).

Answer 2

It sounds simple, but you can simply combine the two vectors into one. Many researchers do this.

Answer 3

What you came to is an important open issue. Yes, there are some ways to handle this, as Eamorr mentioned here. For example, you can combine and execute the PCA method (or the nonlinear dimension method method ). But it is partly difficult to protect the practicality of this, given that the PCA takes O (n ^ 3) times in the number of functions. This in itself may be unreasonable for data in a vision that can have thousands of functions.

Answer 4

As others have mentioned, the easiest approach is to simply combine the two sets of functions into one.

SVM is characterized by a normal to the hyperplane with a maximum margin, where its components determine the weight / importance of features, so higher absolute values ​​have a greater impact on the solution function. Thus, SVM assigns the weights of each function independently.

In order for this to work, obviously, you will have to normalize all the attributes in order to have the same scale (say, convert all functions to the range [-1,1] or [0,1])