Face embeddings

[Berkay Inan]

Hello,

I am master's student at EPFL and I must say that I am very impressed by this dataset and the research. I was interested in implementing a similar system that is presented in the paper, "Looking to Listen at the Cocktail Party".

In the paper, face embeddings are mentioned as,

We use a pretrained face recognition model to extract one face embedding per frame for each of the detected face thumbnails.

We use the lowest layer in the network that is not spatially varying, similar to the one used by Cole et al for synthesizing faces.

Here, which face recognition model is used for extracting 1024D embeddings? Cole et al used Facenet and VGG-Face for their work. Is face recognition model used in the paper one of those, or is it some other novel model?

Best regards and thanks in advance,

Berkay Inan

[or...@google.com]

We actually played with both FaceNet and VGG-Face, and did not see a significant difference between them. In FaceNet, the results were obtained using the avgpool-0 layer, and in VGG-Face they were obtained using fc6 layer.