Data Set Description For Labeled Faces in the Wild (LFW)
There are total 13233 images in the LFW data set. 5749 people with 1680 people with two or more images. The LFW image size is 250x250. All images are collected from the web. Each face has been labeled with the name of the person pictured. Although it claims that these faces were detected by the Viola-Jones face detector. But in our test, we find that there are 11 images that we can't detected by all 4 haarcascade_frontalface.xml files. None of the images are manually labeled with landmarks point which make this data set a unique for testing your face recognition related algorithm. We use LFW data set to testing our ASM/AAM fitting algorithm and also the face match algorithm.
How to do the ASM/AAM Fitting on LFW?
Since the LFW face database is not manually labeled, so there is no training set on our ASM/AAM. We just apply the 68 points of ASM/AAM model from the XM2VTS to the LFW and see how our algorithm works on this real world face database. Our tests show that if the face is frontal and no complicated image background, then the ASM/AAM result is pretty good. Of course, you may notice that for some face the ASM/AAM fitting is bad. There are several reasons for that: face pose, blur image, image background, etc. Nevertheless, we think that we are the first site that apply the ASM/AAM technology to the real world face images.
How to do the face match on LFW?
How to use AAM do the face match is an interesting topic.
Althrough there are some papers talk about using AAM do the face recognition, but we really don't see some good results.
Here is what we did:
Using AAM to extract each face feature which is a low dimension (168 data).
Then use the face feature data to do the mapping. For each face in LFW, find the top 5 high similarity match.
The speed is pretty good; It only cost 400ms to match one face against 13233 faces.
(The loading all 13233 face features data into system will cost 19 sec.)
The match rate is not too bad if the same person's face images are similar enough.
Click the left menu to see the result. You may notice that even for the same person, the images sometime are quite different which is the main reason it failed to match.
Let's do an interesting test:
Two similar images below are taken with same digital camera.
If we put one face on the LFW database and search the other faceamong all 13234 faces.
Guess what? We do find the same person on the 1st match with the similarity 80.79%; time cost only 400ms. Amazing, right?
ASM/AAM Fitting Notation
- ASM1/xxx_x_x.jpg is the ASM fitting result with triangle mask.
- AAM1/xxx_x_x.jpg is the AAM fitting result with outline mask.
- AAM2/xxx_x_x.jpg is the AAM fitting result with triangle mask.
- AAM3/xxx_x_x.jpg is the AAM fitting result with appearance.
- AAM5/xxx_x_x.jpg is the AAM fitting result with appearance on the mean shape with outline.
- AAM7/xxx_x_x.jpg is the AAM fitting result with alignment to the mean shape with outline.
- AAM8/xxx_x_x.jpg is the AAM fitting result outline mask.
- AAM9/xxx_x_x.jpg is the AAM fitting result triangle mask.
