Face_recognition

[cmlyldz]

I am dealing with dnn_face_recognition example it works but I have question about face chip details functions. In this example you use deep learning for face recognition but the algorithm can recognize a face with using 1 training image(face). I want to use 2 different image(posing, or another things) for same person in training. Is that possible ?

[davisking]

If you have multiple faces use a k-nearest-neighbor or linear SVM classifier.

[cmlyldz]

Thanks for the answer.

[Kagami]

If you have multiple faces use a k-nearest-neighbor

Sorry for my ignorance but I searched for all mentions of "k-nearest" in dlib's source code and only found 3 suitable functions:

find_k_nearest_neighbors
find_k_nearest_neighbors_lsh
find_approximate_k_nearest_neighbors

What it does is creates a graph based on distance of provided samples. But if I already have labeled samples and need to find a label for the new sample, how would that help?

Isn't it simpler to just find minimal distance in for loop against all known samples? Sorry if I don't understand something.

For example scikit's KNeighborsClassifier has special predict method.

[davisking]

Yes, use a simple for loop. You don't need to do anything more complex unless the for loop ends up being too slow.

[Kagami]

Thanks, I have one more question. Could svm_multiclass_linear_trainer provide better classification results than search for minimal distance? I mean I can always try it, just interested in your opinion because I'm lacking experience in ML.

[davisking]

It might be better. You have to experiment and see what works best though. But I wouldn't be surprised at all if it gave better results.