Function train_detection_network

examples/dnn_instance_segmentation_train_ex.cpp:191–317

}

det_bnet_type train_detection_network(
    const std::vector<truth_image>& truth_images,
    unsigned int det_minibatch_size
)
{
    const double initial_learning_rate = 0.1;
    const double weight_decay = 0.0001;
    const double momentum = 0.9;
    const double min_detector_window_overlap_iou = 0.65;

    const int target_size = 70;
    const int min_target_size = 30;

    mmod_options options(
        extract_mmod_rect_vectors(truth_images),
        target_size, min_target_size,
        min_detector_window_overlap_iou
    );

    options.overlaps_ignore = test_box_overlap(0.5, 0.9);

    det_bnet_type det_net(options);

    det_net.subnet().layer_details().set_num_filters(options.detector_windows.size());

    dlib::pipe<det_training_sample> data(200);
    auto f = [&data, &truth_images, target_size, min_target_size](time_t seed)
    {
        dlib::rand rnd(time(0) + seed);
        matrix<rgb_pixel> input_image;

        random_cropper cropper;
        cropper.set_seed(time(0));
        cropper.set_chip_dims(350, 350);

        // Usually you want to give the cropper whatever min sizes you passed to the
        // mmod_options constructor, or very slightly smaller sizes, which is what we do here.
        cropper.set_min_object_size(target_size - 2, min_target_size - 2);
        cropper.set_max_rotation_degrees(2);

        det_training_sample temp;

        while (data.is_enabled())
        {
            // Pick a random input image.
            const auto random_index = rnd.get_random_32bit_number() % truth_images.size();
            const auto& truth_image = truth_images[random_index];

            // Load the input image.
            load_image(input_image, truth_image.info.image_filename);

            // Get a random crop of the input.
            const auto mmod_rects = extract_mmod_rects(truth_image.truth_instances);
            cropper(input_image, mmod_rects, temp.input_image, temp.mmod_rects);

            disturb_colors(temp.input_image, rnd);

            // Push the result to be used by the trainer.