Can we recognise half face without 1 eye?

[cat3333]

Hi, thanks for this great jog and answering my last question.

I have run the 'static_input' example with the INDIVIDUAL_FRONTAL_AFLW_SPLIT_1.xml model, and find that the model can detect the faces with 2 eyes but not with only 1 eye. the results are as below.

successful result:

failed result:

I would like to know that

1. can we detect the face with 1 eye using code snippet (http://cmp.felk.cvut.cz/~uricamic/clandmark/#snippets-mv) you provided?
2. If yes, what model we should use in "modelDir" in the main function in the code snippet?

[cat3333]

The photos I used are from http://xdesktopwallpapers.com/wp-content/uploads/2012/04/Zooey%20Deschanel%20Sad%20Side%20Face%20Closeup.jpg http://74211.com/wallpaper/picture_big/Anne-Hathaway-HD-Post-in-Pixel-of-1920x1440-in-Smiling-Facial-Expression-and-Messy-Hair-She-is-Impressive-and-Fit-TV-Movies-Post.jpg and http://mag-cosmo-prod-s3.s3.amazonaws.com/legacy/images/beauty/neck.jpg

[uricamic]

Hi @cat3333,

I think the main problem is that those faces have too big inplane rotation (roll angle), which makes them difficult to be detected by a conventional face detector such as OpenCV haarcascade used in the examples.

CLandmark expects detected face box on its input, so you can use any technique you want to improve the results. Here are the examples using a commercial face detector (implementing this paper), which is able to detect faces in a full yaw range and inplane rotation up to 90 degrees:

So, to answer your questions:

• It is indeed possible to detect close to profile faces (even with a strong inplane rotation), you can use haarcascade for profile faces (and try to rotate the image in order to deal with the inplane rotations).
• The modelDir in the code snippet should contain the folder with multi-view model files (xml files), which are listed in the modelList array.

[cat3333]

Thank you very much for the detailed solution!!!