opencv - SVM classifier testing works only on images bigger than the training set -

opencv - SVM classifier testing works only on images bigger than the training set -

i'm implementing, first time, sw objects detection static images. first goal observe simple circles, i'll move more complex object. unfortunately seems have problem when validating classifier.

my selection utilize pig descriptor (using opencv) , svm classifier (using svmlight). code compiles , works there sounds odd me, concerning svm.

i have:

a training set composed 5 images 48x48px of different circles , 5 images 48x48px of non-circles (i know there few of them in order have solid classifier but, know, it's test works) a test set composed 4 images 48x48px (with circles big ones used training) , 1 image much bigger (765x600px) multiple size circles , other geometric forms.

what happens that:

the circles in test set not detected when images 48x48, if in test set there images used in training phase. in image 765x800 (which contains circles of size) circles of same size of training set, or bigger, correctly identified.

i'm using next parameters:

hog: winsize=48x48px, winstride=4x4px, cellsize=4px, blocksize=8px, blockstride=4x4px classifier: svm regression linear classifier c=0.01. (rbf results worse linear)

this api performs detections parameters i'm using.

vector<rect> found; double hitthreshold = 0.; // tolerance size padding(size(32, 32)); double scale = 1.05; int groupthreshold = 2; hog.detectmultiscale(testimg, found, hitthreshold, win_stride, padding, scale, groupthreshold);

is there reason why circles in images 48x48px not detected , circles in bigger image detected? expect 48x48px images correctly classified in order validate classifier. have added bigger image when nil detected in 48x48px images.

besides, sounds stranger fact in 48x48ps test set there images used in training set , think must identified, instead not! (i know training set , test set must different did when nil detected.)

this first experience pig descriptors , svm might not work because of configuration error or selection of images..

any help welcome!

thanks in advance :)

opencv image-processing svm svmlight