Umap

Hi @cgilles @thanhtrungdinh , This merge request includes the experimentation codes of facenet recognition model and Umap dimension reduction method. The result is outstanding compare to the current Openface model that we are using. The accuracy of facial recognition with outlier detection is 93% when only using facenet and goes up to 97% when having umap to project the data on 2D plane. With this new facenet model, we don't need face alignment anymore but the dimension of the output vector is now 512 (4 times larger than the output of the current openface model). Another point is that it would improve processing time and accuracy greatly with umap, however, we haven't found any suitable C++ implementation of this algorithm.