Hi Angeliki,
> 1) Is it true that the LFW just measures performance from comparing
> images, not classifying people while using FaceNet?
Correct, the LFW just compares whether two faces are of
the same person or not. The performance on this
correlates well with classifying people.
Yes, everything is there, see our setup guide at
http://cmusatyalab.github.io/openface/setup/
I recommend using the Docker image if you just want to
try it out first since the dependencies are difficult
setup at first.
> 3) I will need to test the algorithm with a new dataset in order to make
> my conclusions. Do I have to retrain the CNN with my dataset (if yes, in
> which way)?
If you want to classify people, you don't need to retrain the CNN, see:
http://cmusatyalab.github.io/openface/demo-3-classifier/
If you have a lot of data and want to try to re-train the CNN, see:
http://cmusatyalab.github.io/openface/training-new-models/
> 4) Is there any official online guide to help me understand how do train
> and use Facenet?
See their paper:
http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Schroff_FaceNet_A_Unified_2015_CVPR_paper.pdf
And also take a look at our comparison demo:
http://cmusatyalab.github.io/openface/demo-2-comparison/
We also have API docs online at:
http://openface-api.readthedocs.org/en/latest/index.html
I'm not familiar with this.
> 6) Do you know if FaceNet can detect and recognize faces that are not
> frontal and for example only one eye is obvious at the image?
It won't work well for side-profile faces.
You can get a better idea for what kind of faces it works for
by using our web-based demo:
http://cmusatyalab.github.io/openface/demo-1-web/
> 7) I forgot to mention that I have windows 7 and that I don't I really know if
> I need Torch for running FaceNet; and if yes can I have it at windows?
OpenFace needs Torch, Python, opencv and dlib.
They should all work on Windows, but I only use the code in
Linux and OSX and there will probably be some cross-platform
issues you'll need to fix. I'd be happy to take a PR
fixing them for future users.
-Brandon.