Compare different face images

[yun-long]

Hi,

I am trying to separate many face images into different classes. e.g., identify all faces of a person, named "Sarah", and put all identified face images into one folder.

Firstly, I download this model (20170511-185253) from this repository. Secondly, I forward all my face images into this model and save the embedding of faces. Thirdly, I calculate the L2 distances with the embedding, between unknown faces and my reference face "Sarah".

Finally, I got this result shown in below figure. The first image is my reference image, "Sarah", the numbers are the distances, the images with red border are identified "Sarah".

But, the problem is that, the distances, d1 = "reference Sarah" - "real Sarah", d2 = "reference Sarah" - "not Sarah" are quite the same.

Does anyone known how can I improve it?

[huan]

+1

That's the same problem I'm facing to...

[davidsandberg]

Hi, What input pipeline did you use, i.e. image standardization, alignment, etc?

[yun-long]

@davidsandberg hello, the pipeline looks like this:

some image => face detection => crop face => pre whiten => neural network => face representations => euclidean distance

I did't use face alignment. Should I use it? Thanks

[shangwenxiang]

@davidsandberg +1,That's the same problem I'm facing to...

[GuiUzeda]

I'm having the same problem. The euclidean distance seams to be very low for all the faces compared to the facenet paper. In my case, I used face align with dlib

A provisional workaround for me was to lower the threshold to a point where I didn't get false positives anymore. I can see that your code might have a threshold equal to 0.04. If you lower that to 0.03 you will probably eliminate the false positive, but you would have more false negative.

+1 to that problem, could not solve this particularly well yet

[axmadjon]

@davidsandberg

At me too such problem. how can you solve this problem? Who knows?

[rexlow]

+1 Any innovative ideas yet?