LBPHFaceRecognizer model

asked 2020-04-22 10:52:41 -0600

updated 2020-04-23 02:32:39 -0600

berak
32993 ●7 ●81 ●312

Hello everyone,

I am trying to train a LBPHFaceRecognizer model with photos from a lot of persons. The problem is that at a point I get an OutOfMemory error. I looked up at the model after I saved in a file and it seems to be saving the histograms of every photo and i think this is the problem. From my point of view it should be saving a representative set of features for every person (If I have 50 photos with the same label it should be saving 1 representative histogram (maybe the centroid of the histograms cluster), not 50). Maybe I am wrong, but if is there someone who knows better what happens in the training and predict phases for this algorithm or someone who used it for a big dataset, please reply.

Thank you, Bogdan

edit retag flag offensive close merge delete

Comments

it seems to be saving the histograms of every photo

this is correct (and this is all, that happens in the training phase, the prediction is just a 1-nearest neighbour search over that)

maybe the centroid of the histograms cluster

i don't think this is feasible (but please try and report back !)

from a lot of persons.

how many ? again, since it's a linear search (and not building a "global model"), you could split it up into several instances

berak ( 2020-04-23 02:42:07 -0600 )edit

Thank you for the reply. I am trying to train the model on a dataset that contains about 1500 persons and a like 50-100 photos for each person. I guess using so many photos with this LBPHFaceRecognizer is not the best approach.

Bogdan133 ( 2020-04-25 07:31:44 -0600 )edit

50-100 photos for each person.

well you certainly need a few. maybe you should do some cross-fold validation to find out the minimum number needed.

maths time: 64 x 256 x 4 = 64kb per image. for 1500 persons and 50 imgs each, we're at 5gb mem, that's quite a lot.

berak ( 2020-04-25 07:43:37 -0600 )edit

Try training neural network (cnn)

holger ( 2020-04-26 20:51:24 -0600 )edit

add a comment

answered 2020-04-25 08:04:01 -0600

berak
32993 ●7 ●81 ●312

you could try to use different, shorter features instead of lbph.

opencv's dnn supports using the openface model, which makes a 128 float feature from an (bgr) image:

// https://storage.cmusatyalab.org/openface-models/nn4.small2.v1.t7
dnn::Net  net = readNet("nn4.small2.v1.t7");

// for each image:
Mat inputBlob = dnn::blobFromImage(img, 1./255, Size(96,112), Scalar(), true, false);
net.setInput(inputBlob);
Mat res = net.forward().clone();

// now cache the result Mat's in some database, and do your own
// nearest neighbour search using e.g. 

// L2 norm:
float dist = norm(a,b); 

// or a simple dot:
float dist = a.dot(b);

// or cosine distance:
float x = a.dot(b);
float y = a.dot(a);
float z = b.dot(b);
float dist = - x / sqrt(y*z);

edit flag offensive delete link

add a comment

LBPHFaceRecognizer model

Comments

1 answer

Links

Question Tools

Stats

Related questions

LBPHFaceRecognizer model edit

Comments

1 answer

Links

Question Tools

Stats

Related questions

LBPHFaceRecognizer model