Deep funneling in FaceNet about facenet

[davidsandberg]

Hi,
The model you are referring to has been trained on a combination of CASIA-WebFace and FaceScrub datasets aligned using the Dlib face alignment (and align_dataset.py). No deep funneled images has been used for the training set. With the Dlib face alignment quite a few faces are missed and these images will not be included in the training set. For the CASIA-WebFace dataset i think this is around 50 000 images out of 490 000. This effectively filters out some of the difficult examples from the training set, like faces with sun glasses and other partially occluded images. My feeling is that this makes the training set a bit too "easy".
I have tried a couple of alternative face alignment algorithms, like Clandmark but this one missed more faces than Dlib on the LFW dataset. Currently I'm trying TCDCN which looks promising but I haven't used it for training yet. And here you can find a nice list of face alignment packages.

[Jusya]

You helped me a lot to understand how this face recognition system works)
I'm very thankful to you!
I'm going to play with different face alignments you've suggested)