trying to make face recognition with ANN

edit

opencv comes with a MLP readymade that you could use.

here is a nice post on how to prepare the data (for an svm, but that step is the same for ANN, too)

apart from that: if you're really into ann's allow some time for this

that post is realy good:) i already made data loading, now just reading the "MLP readymade" to get a grip how its made :)

found an mlp example

can someone explain me why i get memory allocation problem?

int sks[4]={10,10,5,1};
  CvMat skuoksniai= cvMat(1,4,CV_8UC1,sks); // not sure about CV_8UC1 part as well :/
  CvANN_MLP tinklas;
  tinklas.create( &skuoksniai,CvANN_MLP::SIGMOID_SYM,1,1); //problem line.

int sks[4]={10,10,5,1};
cv::Mat skuoksniai= cv::Mat(1,4,CV_32SC1,sks); // just use CV_32S
CvANN_MLP tinklas;
tinklas.create(skuoksniai,CvANN_MLP::SIGMOID_SYM,1,1); //problem no more i hope ;)

(yea, i know, that cv::Mat, cvMat is annoying ...)

is that lithuanian above ;) ?

thank you for quick response. k, ill leave CV_32SC1 part for later. but when i remove & i get "no instance of overloaded function "CvANN_MLP::create"" :( yes :)

damn, works fine here ...

dude, sry my bad.. didn't sow the "cv::Mat" part. now its seams to run :) thank you so much :)

and one again i get memory allocation problem. cv::Mat trainData(17,38416,CV_8UC1)// its 17 rows and 38416 columns of grey images 1 row = image
cv::Mat label= cv::Mat(1,17,CV_32S,temp);// temp is int temp[17]; witch contains {0,0,0,0,0,0,0,1,1,1,1,1,2,2,2,2,2} cv::Mat weights= cv::Mat::ones(1,17,CV_32S); tinklas.train(trainData,label,weights);

i tried cv::Mat new; cvtColor(trainData,new,CV_32S);

http://docs.opencv.org/modules/ml/doc/neural_networks.html#cvann-mlp-train

• your trainData Mat has the correct shape, but the wrong type. you need float data here, so it needs a simple conversion, before you apply it: trainData.convertTo(trainData, CV_32F); (cvtColor is more for translating from bgr to hsv or such, it does not change the datatype)
• your labels have the wrong shape. for 17 feature vecs(rows), you need 17 labels(rows) as well, so: cv::Mat label= cv::Mat(17,1,CV_32S,temp); // only rows/cols swapped to what you got before

again, don't worry too much. you're doing fine. all that confusion is pretty normal ;)

yeah, i did a bit reading bout CV_32F c1 c2 and etc :) ill gona try thank you again :)

still same error, ill gonna continue tomorrow, and ill post if find anything :) thank you for your help, its getting bit clearer :D

ok, so found reason for that error, it was my silly mistake, forgot to set correct first layer. now i have another error, is with prediction. so the label is the same

cv::Mat label= cv::Mat(17,1,CV_32F,temp); cv::Mat pazinti(0,38416,CV_32F); tinklas.predict(pazinti,label);

and i get this error: OpenCV Error: Assertion failed (dims <= 2 && data && (unsigned)i0 < (unsigned)si ze.p[0] && (unsigned)(i1DataType<_Tp>::channels) < (unsigned)(size.p[1]channel s()) && ((((sizeof(size_t)<<28)|0x8442211) >> ((DataType<_Tp>::depth) & ((1 << 3 ) - 1))*4) & 15) == elemSize1()) in cv::Mat::at, file c:\opencv\build\include\op encv2\core\mat.hpp, line 537

morning ;)

cv::Mat pazinti(1,38416,CV_32F); // 1 feature vector for testing (1 row)
cv::Mat label(1, 1, CV_32FC1 ); // as many rows as your output layer (1 here?)

btw, it might be a bit early for that, but to my experience, ann works better with 1 output per class. (you look for the maximum output, that's the actual class prediction)

morning :) well yes my output layer is 1. but i thought label supposed to be the same as in training, now i assume that the predict value will be stored in there?

" i thought label supposed to be the same as in training" - you train with 17 features and 17 resp. labels, and predict with 1 feature and num_outputs labels (a here, again[so far..])

"now i assume that the predict value will be stored in there?" - yep. float cls = mat.at<float>(0,0);

ok, so in that case everything should work, but im geting "OpenCV Error: Assertion failed (layer_sizes != 0) in CvANN_MLP::predict, file ........\opencv\modules\ml\src\ann_mlp.cpp, line 1611" could it be that i have created and trained ANN in a void function? i guess no, because i can use .predict

structure main{ void{load images, create/train ANN} void{webcam stream, .predict} }

"could it be that i have created and trained ANN in a void function?" -hmm, no idea what you mean.

if the problem does not resolve on its own, you might need to put your code on a pastebin, and sent the link here..

how as i thought, i have main function which Creates ANN and trains it (1 time), and i have another function which is in a loop and i want that in there i could use .predict(). so my question: what parameters i should send to my function (which is in a loop) that it could use .predict() ? or maybe i could make ANN global or smth?

oh, is it about passing arguments now ?

int myPredict( CvANN_MLP & ann ) { /* do the prediction*/ } int main() { CvANN_MLP ann; ann.train(...); int p = myPredict(ann); }

on the other hand, you usually train such a beast only once. then you save() it to xml or such, and the next time load() that model instead of retraining.

honestly, i can only guess, what you mean there. ;(

thank you once again :) i will make that ** work :D in this point you deserve a medal for helping me out ;)

ok, i guess its working kinda :D i get from .predict(input,label) data i get it by cout<<(int)label.at<float>(0,0) but its always the same result, no matter what is input. i guess i need to edit ANN.train(). btw i was assuming that it would be trained as backprop and not as rprop.

oh, cool ;)

yea, the fine-tuning will take some work. also, criteria.max_iter, criteria.epsilon, etc.

how do your skuoksniai look like now ? (i love that word..)

lol, i just now sow that i made a mistake in that word :D it should be sluoksniai :D nwm layers = 38416, 300,100,50,1 its random values actualy. but im quite concerned about that CV_32F part, because when i use imshow() the image is almost all white, i wonder does it affect pixel values. and when i use ANN.predict(input.label) (like 10 times with different input) the output doesn't change :(

google thought so, too. ;9

what are you trying to imshow() ? if it's a float img, if it's not in the [0..1] range, it might need some scaling, like img/255 or similar.

also, if you're training with backprop, you'll probably need a few ( thousand ;] ) iterations

yeah, google knows it all :D i checked that transormation, everything is alight, i guess now just need to make backprop with parameters :) could you explain in human language what sampleIdx should contain? talking about CvANN_MLP::train