Emotions from profile faces

asked 2015-04-29 08:35:26 -0500

amit shah gravatar image

I am currently workin on profile faces. I want to be able to detect faces with side views. Is there a way other than having to use haarcascade_profileface.xml ? This cascade tends to have large false alarm rate. Besides, I want to learn some other technique to deal with it. The Facebook-tag people seems to be very much robust while detecting people with side views. Any algorithm which can achieve such robustness? My ultimate goal is to be able to determine emotions from side viewing faces if at all thats possible.

edit retag flag offensive close merge delete

Comments

4

"ultimate goal is to be able to determine emotions from side viewing faces" - you will need a total different approach for this. cascades (of any kind) are for detection, yours is a classification task.

(i'd rather go and train a facial landmarks detector (like dlib's) on profiles, and use an SVM for classification)

btw, you can handle the false alarm rate by increasing the minNeighbours param.

also, try: https://github.com/nenadmarkus/pico (it's incredibly fast, like 5ms for 640x480)

berak gravatar imageberak ( 2015-04-29 08:41:54 -0500 )edit

Nice link!

StevenPuttemans gravatar imageStevenPuttemans ( 2015-04-30 02:43:44 -0500 )edit

Does SVM training for facial landmark detection involve annotating the images(marking the images with landmark points) and feeding this data to SVM ? How many images ideally are enough for robust facial landmarking? How would the SVm behave if the images I used to train SVM where exactky half side faced and the testing image is more covered than the images used for training?

amit shah gravatar imageamit shah ( 2015-04-30 08:53:53 -0500 )edit
1
  • "Does SVM training for facial landmark detection involve annotating the images" - well, yes. but you'd rather train an automatic landmark detector on images, than manually annotating 100'eds of images. (you'd need that detector later anyway, for the classification after successful training)
  • "How many images" - a few hundred, probably
  • http://blog.dlib.net/2014_08_01_archi... (real shame, that opencv does not have something similar)
  • " How would the SVm behave" - pardon ? can you explain "and the testing image is more covered" ?
berak gravatar imageberak ( 2015-04-30 09:04:28 -0500 )edit

@berak, do you know if "pico" is patented?

Gino Strato gravatar imageGino Strato ( 2015-05-01 02:07:44 -0500 )edit

@Gino i don't know..

the code is copyrighted for sure, but i don't think, there's an actual patent on the idea

berak gravatar imageberak ( 2015-05-01 02:09:58 -0500 )edit

@berak, I've tested the face detector and it seems great! I've looked for some information about the algorithm, but I've mainly found links to their own website. Furthermore, incredibly there are only 2 citations reported by google scholar. How did you know about it? Anyway, I think that a self-implementation code of the ideas reported on the paper are not prohibited.

Gino Strato gravatar imageGino Strato ( 2015-05-01 02:44:08 -0500 )edit
1
  • "How did you know about it? " - found the paper somewhere
  • "self-implementation code of the ideas reported on the paper are not prohibited." - yes, i'd think so, too ;) still, why not just ask the author ?
berak gravatar imageberak ( 2015-05-01 02:47:54 -0500 )edit
1

I've just sent an email to him.

Gino Strato gravatar imageGino Strato ( 2015-05-01 14:23:47 -0500 )edit

@berak: "you'd rather train an automatic landmark detector on images". How can that be achieved?

amit shah gravatar imageamit shah ( 2015-05-04 00:53:06 -0500 )edit