4 ms·
I'm not sure how leading edge the openbr stuff is, AFAICT (i'm not a domain expert or even novice), it's pretty primitive. If you were starting from scratch tod
by asteadman 10y ago
I'm not sure how leading edge the openbr stuff is, AFAICT (i'm not a domain expert or even novice), it's pretty primitive. If you were starting from scratch today, i'd say the openface project (https://cmusatyalab.github.io/openface/ https://cmusatyalab.github.io/openface/) is a better starting point.
Having a DNN that maps images of faces into a point on a hypersphere seems eminently practical. I'm a bit surprised that it works, but having something like that is awesome.
- tyingq 10y agoThis post seems to indicate they aren't that far apart: http://bamos.github.io/2016/01/19/openface-0.2.0/ http://bamos.github.io/2016/01/19/openface-0.2.0/ I am though, in the same boat as you in this domain...not even a novice.
- asteadman 10y agoThat post is great! thanks for the link, i hadn't stumbled across it yet. From what I can tell from that link though, it is exactly as I suspected, and openface 0.2 significantly outperforms openbr for the LFW test set. (whereas openfaces 0.1 was largely equivalent). Moreover, I think that there is clearly room for drastic improvement with the openface architecture, either by tweaking the layout of the NN's, or simply throwing more data at it. It is not clear to me how openbr could be improved.