Facenet: Classifier results are not correct

[allahbaksh]

Hi @davidsandberg
We trained the classifier on the LFW dataset. We picked the pre-trained model from the GIT Repo. When we try to classify the results are poor. If we use SVM with Linear classifier the prediction was 10% accurate and when we tried using RBF then the result was 25%. Whether I am missing something?

[MaartenBloemen]

This is a problem with scikit SVC. Basicly your model is horribly underfitted. SVC multiclass classification uses "One-vs.-one" strategy, blunty it makes K(K-1)/2 hyperplanes where K is the number of classes. For LFW this means there will be 16522626 hyperplanes with only 13233 data points, starting to see the problem here?

[allahbaksh]

@MaartenBloemen Will kNN solve this problem? Have tried any other classifier for this.

[MaartenBloemen]

Ik don't know if kNN will perform better I haven't used it yet. The approach I took in my project was how SVM was intended as a binary classifier. Each person has it's own classifier containing the person and "unknow". Then to be able to go through alot of classifiers in a short time I made a distributed system where each of my worker nodes is assigned one or more classifiers. When I need to classify a face in a frame I simply send the embedding Array to all my workers and wait foto a response.

But of you try kNN and it performer good be sure to let me know! :)