如何使用SIFT功能/描述符作为SVM培训的输入？

Question

I want to classify MRI images of a brain tumor into benign and malignant using C++. I am using SIFT features and the paper I am following clustered them using kmeans before training the SVM classifier. What I don't understand is why is there a need to do that? From what I know, kmeans only clusters the features; it doesn't change the size of the input.

I have read that possible ways are BoW and histogram. In the histogram approach , It just counts the # of features in each cluster right? I don't think that will provide the information I'll need for classifying benign and malignant tumors because they can be both small and big. In BoW approach , I didn't understand this link .

Basically, I don't know what to do with my SIFT features to use it as input for SVM. Do I really have to create a dictionary of some sort? I'm begging you, please enlighten me. Thank you very much!

Answer 1

I'm not too familiar with OpenCV or SIFT features, but this should be general enough to be useful to all programming languages. I will also be describing only the BoW approach below.

Let's assume we had N images. For each image i , we have F number of features, and each feature had D dimensions. We can put all the features into an array feats , so that it looks like this:

[1, 2, ..., D]
[..., ..., ..., D]
[N*F, ..., ..., D] 

Each row of feats is a feature, with D dimensions, and we have a total of N*F features.

In k-means, we take all these features and group them into k clusters. Therefore, every single feature is assigned to a single cluster. Most k-means functions typically return a matrix C of size kx D , which represents the centroids of the clusters. This matrix C is the "codebook" or "dictionary" of the k-means algorithm. Some also return a vector of size N*F which shows which cluster each feature is assigned to (in OpenCv, this is represented by the labels variable in this link: http://www.developerstation.org/2012/01/kmeans-clustering-in-opencv-with-c.html ).

Since we already have the assignments of all the features, each image i has F features, which can be simply represented by the clusters they belong to. For example, if the original image was represented as

[1, 2, ..., D]
[..., ..., ..., D]
[F, ..., ..., D] 

then the image can also be represented simply as a vector:

[1] % Assignment of feature 1
[...]
[F] % Assignment of feature F

Therefore, you can take this vector and form a histogram h of the clusters that are represented. This histogram is the feature vector for the image, which you can later use in the SVM.

PS If you need any further clarification and/or an example, let me know!