|
For example I have N subsets of croped human faces, each subset have images of the same person, but subsets are different in size but >=2 images.I think first problem is that croped faces must be preprocessed and normalized in some way. Or I have N face pairs that people consider similar. So I want to learn some "metric" or "function" of similarity from this data to compare other unknown faces for similarity. In other words I want to find some high level features from raw pixels. Any suggestions/recomendations? |
|
The baseline features for this task is extracting EigenFaces (PCA on the raw pixel vectors) from your faces corpus and projecting onto it, see for instance this example in scikit-learn. Off-course the state of the art is to use much more complicated models (e.g. 3D models of faces using deformation features with a lot of hand-made task-specific features or maybe automated non linear feature extraction using convolutional deep belief networks for instance). Hmm, maybe task is can be treated as projection to some nonlinear subspace and than we can use simple euclidean distance, but how to take into account that (as I think) we should have small distance in group of faces(of one person) and big distance between face groups(different persons).So we need to find this subspace somehow.
(Feb 01 '13 at 06:45)
mrgloom
Also I found this http://graphics.cs.cmu.edu/projects/crossDomainMatching/ but I don't know is it suitable for faces.
(Feb 01 '13 at 06:48)
mrgloom
Have a look at DrLim by Lecun. It finds a metric based on "similar" and "dissimilar" pairs.
(Feb 01 '13 at 08:00)
Justin Bayer
thanks, very intersting paper http://www.cs.nyu.edu/~yann/talks/lecun-20061208-nips-similarity.pdf
(Feb 04 '13 at 04:02)
mrgloom
|
