DNN forwards only first image

edit savecancel

opencv version ? what kind of network do you load ?

OpenCV 4.1. I use CNN with 64x64x3 on input and sigmoid on output layer. I got it by converting keras→tensorflow→openCV DNN. It works correctly on single image.

how should the output look like ? what is expected ? (is it just a classification network ?)

It is a binary classificator. When I run it on single images i got correct outputs.

cv_pred = []
for img in imgs:
    cv_input = cv.dnn.blobFromImage(img)#, size = img.shape, swapRB=True, crop=False)
    cvNet.setInput(cv_input)
    cvOut = cvNet.forward()
    cv_pred.append(cvOut[0][0])

Out:

[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0]

But with running in multiple images I got only one prediction: array([[1.]], dtype=float32)

seems to work as expected with 4.2.0 or master. can you try to update ?

The same result on python with cv__version__ = 4.2.0

Did I correctly understand? You got multiple output after blobFromImages and one forward command?

unfortunately, yes, it works for me

btw, vector<Mat> pred; not vector vector (and that works, too ;()

try :

nom_modele ="bvlc_googlenet.caffemodel"
nom_proto = "bvlc_googlenet.prototxt"
nom_classe = "classification_classes_ILSVRC2012.txt"
caffe = cv.dnn.readNet(nom_modele, nom_proto)
try:
    with open(nom_classe, 'rt') as f:
        classes = f.read().split('\n')
except:
    classes = None

img_original1 = cv.imread(r"C:/tmp/neuralstyle/elephant.jpg")
img_original2 = cv.imread(r"C:/tmp/neuralstyle/zebra.jpg")
if img_original1 is not None and img_original2 is not None:
    moy_image = cv.mean(img_original1) 
    img_all = np.array([img_original1,img_original2])
    blob_all = cv.dnn.blobFromImages(img_all, 1.0, (227, 227), moy_image, swapRB=False)
    caffe.setInput(blob_all)
    val_sorties = caffe.forward()

I tried. But it doesnt work.

I see that my original keras model uses «channels last» so input array has shape (20, 64, 64, 3) But opencv works with (20, 3, 64, 64). Could this fact be causing the problem? I guess no but ..