We are going to add landmarks information in one of the next releases. I think It could be calculated from this data. But I understand that it's better to add such functionality to CompreFace, I'll add this to our backlog

About comparison with other systems - as I mentioned we use Facenet under the hood. According to their GitHub, the accuracy on LFW data set is 99.65%: https://github.com/davidsandberg/facenet#pre-trained-models

It sounded like the previous comment wasn't asking for orientation detection as a feature. I interpreted the question to be asking under what orientations CompreFace can be expected to work.

Yes, what are the allowed degrees of rotation away from looking at the camera that a face will be detected? And is there a curve demonstrating the FR accuracy as face rotation away from looking directly at the camera increases?

hmm... good question. I haven't found any information for FaceNet that we use. Do you have similar information for any other face recognition system? I would like to see how they calculate it.

First you populate the gallery with the recovered orientation of each face, so the gallery can be segmented by pose, and then the laborious process of humans evaluating the accuracy via generation of a ROC Curve for the gallery. Which can be segmented by face orientation to show the accuracy as orientation varies. People generally hate calculating these because it requires a human to evaluate the FR matches, and mental fatigue set in very quickly.

have you considered to switch to https://github.com/deepinsight/insightface as base model?

edit: never mind, you do https://github.com/exadel-inc/CompreFace/blob/master/embeddi...

Yes, we will publish docker-compose configs that uses insightface with some of the nearest releases. We'll still support Facenet as well, so we'll see what library is more popular.