David Arcus

Hi Davis - I started again and re-trained the network using the default settings in your dnn_mmod example, except I set it to 8000 iterations, and ended up with training and test results very close to 1 across the board. The results are amazing! Thanks very much for your help, and thanks for all the effort you put in to documenting and supporting dlib, I've learned loads. Cheers

Hi Davis - I started again and re-trained the network using the default settings in your dnn_mmod example, except I set it to 8000 iterations, and ended up with training and test results very close to 1 across the board. The results are amazing! Thanks very much for your help, and my special thanks for all the effort you put in to documenting and supporting dlib, I've learned loads! Cheers

Ok, I've made a small detector app that uses the pyramid_up function as follows: while(img.size() < 1800*1800) pyramid_up(img); However none of the cats (from the training dataset) are being recognised. I also tried re-training using all of the defaults in the dnn_mmod_ex example (with only 300 iterations and an average loss of 1.3187) but I still got 0 recall and AP. I'm feeling a bit stumped. Is there something else I could be doing wrong? Thanks again

Thanks Davis. I do believe the cat faces are big enough but I did try training my data using the default settings found in the dnn_mmod_ex.cpp example (with a 40x40 window and smaller network) but I had exactly the same problem. I don't suppose this line could be the problem could it? dnn_trainer<net_type> trainer(net, sgd(0.0005,0.9), {0,1,2,3,4,5,6,7}); I'm getting a huge speed boost using all these GPUs but I'm wondering if it's affecting the results at all.

Hi Davis, thanks so much for coming back to me. My average loss was 0.00146525 at the end of training, with about 50k steps

Hi everyone, I'm trying to train a frontal cat face detector using dlib's dnn mmod tools but am having trouble with the training. I was hoping someone could steer me in the right direction. Using the dnn_mmod example and the find_cars examples I've created a larger network than the default, set the sliding window to 80x80 and set the cropper to its default settings. The trainer is set to 8000 iterations, as per the face detection example. I've labelled about 4,000 cat faces with a square aspect ratio...

Thanks very much Davis. I was able to get it to a matrix using the following: ... while(!win.is_closed()) { cap.set(CV_CAP_PROP_FRAME_WIDTH,200); cap.set(CV_CAP_PROP_FRAME_HEIGHT,150); cv::Mat img; cap >> img; cv_image<bgr_pixel> temp(img); matrix<rgb_pixel> spimg; assign_image(spimg, temp); cv::Mat img1 = toMat(spimg); auto dets = net(spimg); ... Thanks again for all your work on dlib!

Hello, I'm attempting to modify the dnn_mmod_dog_hipsterizer.cpp example to detect dog faces in an OpenCV video. Can someone explain how I can modify the video stream so the detector can be used on it? This is what I've done: ... template <long num_filters, typename SUBNET> using con5d = con<num_filters,5,5,2,2,SUBNET>; template <long num_filters, typename SUBNET> using con5 = con<num_filters,5,5,1,1,SUBNET>; template <typename SUBNET> using downsampler = relu<affine<con5d<32, relu<affine<con5d<32,...