opencv :: dnn中的神经网络输入大小

Question

I'd like to determine (dynamically) the image size expected as input of a deep neural network model that is loaded with the dnn module of opencv. For instance, if I load a caffe model, I first have:

cv::Mat img = imread("my_image.jpg");
cv::dnn::Net m_net;
m_net = cv::dnn::readNetFromCaffe("my_prototxt","my_model");

int img_size; // TODO how to determine it ?

cv::Mat blob =  cv::dnn::blobFromImage(img, 1.0, cv::Size(img_size, img_size));
m_net.setInput(blob);
cv::Mat res = m_net.forward("layer_name");

My problem is to determine img_size from m_net here. I guess it should be feasible with dnn::getLayersShapes but I do not understand what is supposed to be the first argument in my case.

I give here the example of a caffe model, but I am interested in being able to do the same with a tensorflow, darknet or torch model as well.

Answer 1

If you open the ".prototext" file that comes with the caffe model, it states the required input sizes. (Haven't used models other than caffe ones)

Setting correct input size is the responsibility of the user, as some network architectures, eg. fully-connected feedforward networks require exact size inputs, while architectures like CNN's do not and can work on arbitrary sizes of input images.