How to copy vector to a row of a Mat using openCV 2.4.11?

edit

p.s. - why on earth are you using the outdated 2.4 branch ?

something seems to be wrong with this approach: It still fils the Mat column per column. I tried to fill the Mat differently for debugging purposes (and to avoid having to transpose):

     std::cerr<<"der size: "<<ders.size()<<std::endl;
hogFeat.push_back( Mat(ders, true).reshape(1,1) );
ders.clear();
std::cerr<<"hogFeat size: "<<hogFeat.size();

    corresponding output:

     der size: 22680

     hogFeat size: [22680 x 1]read 1

     der size: 22680

     hogFeat size: [22680 x 2]read 2

     der size: 22680

     hogFeat size: [22680 x 3]read 3

     ....

     der size: 22680

     hogFeat size: [22680 x 6]

I would have expected 1x22680; 2x22680 as each row contains the features for ML.

I am filling the Mat in a loop. So what happens is that every iteration I read an image, resize it to 192x192 and calculate the histogram of gradient of it using a 62x62 kernel. I obtain a vector (ders) of length 22680 and am trying to put that into the Mat for ml (SVM), so that later I can perform object recognition. those are the positive samples: http://imgur.com/a/rXK8o and negative ones: http://imgur.com/a/BFDUR (which are supposed to represent the environment of where the people are whose hands we are going to detect). Currently for testing purposes I only used 10 images. The goal is to detect peoples hands in a room to later track them and analyze their movements. Current issue is that no hand is being detected on this image: https://s3.envato.com/files/188980026...

I would have expected 1x22680;

no, it simply prints [W x H], so all ok, i'd think ;)

Current issue is that no hand is being detected

your hog features are very long (10x of the usual), and you're onlyusing 10 images (x1000 needed)

What do you suggest for the hog features, if you say they are 10x bigger than the usual? This is how I did it: https://pastebin.com/w8WVmjmV I think the way i did lays in line with the examples provided by the docs and the infor I could find online.

Also I currently don't have the time to download 1000 images of hands :) , I ll try to do that later, currently I am using this just for testing purposes. Do I understand this corrrectly: if I have 1k images or less I can use eg SVM, while if I have more than 1k images I m better off using neural nets?

the params in hog.compute() influence the size, see here

if you want to train the HOGDescriptor for detection, you''ll also need different SVM params (SVR & LINEAR kernel), again, please look at train_HOG.cpp

... and you simply need a LOT of images in any case, no matter what algo you use.

thanks for the hints, I have read the docs and had a look at the source code. Note that I already have a linear kernel (line 83) and use C_SVC as a SVM type rather than SVR because I want classification (hand/ no hand) rather than regression in order to detect a hand in an image. Feel free to correct me if I misunderstood what you meant or something else.

ah, ok. difference is probably only important, if you want to train it for detectMultiscale()

hmm well I want to detect the location of hands in an image. So I guess I'll need the function detectMultiScale at one moment or another.... So does that mean I have to modify my parameters (although -without wanting to sound rude- what you suggest doesn't seem to be what I need)?

oh, no problem at all.

let's see, what you find, when you're there ?