Image Reshaping with SVM doesn't work about android-face-recognition-with-deep-learning-test-framework

[sladomic]

HI @mkosik

When trying to run a "Recognition Training" for Image Reshaping with SVM, it stucks on a white screen (Training Activity) & it doesn't process any image successfully.

With the current code, it's inefficient to use a high N for image reshaping. For the neural network of facenet we use 160 x 160 as input and get vector of 1 x 128 as output. But if you use the same number for image reshaping, you get 160 x 160 x 4 (channels) = 102'400

Here's the inefficient bottleneck in the SupportVectorMachine class

public String getSvmString(Mat img){
        img = getFeatureVector(img);
        String result = "";
        for (int i=0; i<img.cols(); i++){
            result = result + " " + i + ":" + img.get(0,i)[0];
        }
        return result;
    }

[sladomic]

If you use N < 50, it works. We mostly used N = 25, because the results were similar to N = 50, but it was a lot faster.

[tomsang]

I have a small patch to fix this issue. The performance of Reshaping with N=160 will be close to that of TensorFlow with SVM.

But I don't know how to submit it.

[chithrabaskar]

by changing the package name of this app it did not work...why??

[sladomic]

@tomsang to submit your code you could 1. fork the repo, 2. implement the changes in the forked repo, 3. Create a pull request. I can then test and check it and then merge into the core repo.

[sladomic]

@chithrabaskar does Android Studio give you any errors? At which point does it crash?

[chithrabaskar]

No, there is no error by changing the package name...but it unfortunately stopped while running...please give me any solution

[tomsang]

I would like to provide a improve patch in the later