Automatic Assignment of Tag Icons

[Nghia Duong]

Hello Gilles,

I am planning to, but I am not familiar with the syntax of Atom blog so it
takes me a little much longer to finish. Furthermore, I am at the end of
the semester, where many school projects due in the next 2 weeks, therefore
there is not much time left for me to fill the report. However, I keep
track of the progress in my project journal and I will publish it with the
blog post.

A quick resumé of my progress. I have isolated the code of facial
recognition with the rest of the faces engine, in order to test and
evaluate this module alone. I implement many methods of label assignment
and test them with YaleB extended dataset, with over 11000 face images. By
far I have reimplemented L2 distance, cosine distance, and mean cosine
distance (currently used by faces engine) comparison, Support vector
machine, and K Nearest neighbor for face classification.

The mean cosine distance currently used by faces engine performs well on
small datasets. However, when the data spread, say when there are many
photos of a same face, the accuracy of this method falls to 0%. Currently,
the most promising methods are the closest L2 distance, Cosine distance,
and K-nearest neighbor. Both give an accuracy of around 75% on YaleB
extended dataset and 70% on Lfw dataset. I follow the instructions given by
OpenFace project but it seems to be stuck around this limit.

As a suggestion of Thanh, I am trying to apply UMAP method to find the
general structure of face embedding vectors and plot it for debugging and
clustering. However, if the neural network forward produces too many
outliners, which is incorrect according to the paper, clustering could be
very difficult to give a better result.

Nghia
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-devel/attachments/20200607/eca4eac4/attachment.htm>