<html><!-- Created using the cpp_pretty_printer from the dlib C++ library. See http://dlib.net for updates. --><head><title>dlib C++ Library - dnn_mmod_ex.cpp</title></head><body bgcolor='white'><pre>
<font color='#009900'>// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
</font><font color='#009900'>/*
This example shows how to train a CNN based object detector using dlib's
loss_mmod loss layer. This loss layer implements the Max-Margin Object
Detection loss as described in the paper:
Max-Margin Object Detection by Davis E. King (http://arxiv.org/abs/1502.00046).
This is the same loss used by the popular SVM+HOG object detector in dlib
(see <a href="fhog_object_detector_ex.cpp.html">fhog_object_detector_ex.cpp</a>) except here we replace the HOG features
with a CNN and train the entire detector end-to-end. This allows us to make
much more powerful detectors.
It would be a good idea to become familiar with dlib's DNN tooling before
reading this example. So you should read <a href="dnn_introduction_ex.cpp.html">dnn_introduction_ex.cpp</a> and
<a href="dnn_introduction2_ex.cpp.html">dnn_introduction2_ex.cpp</a> before reading this example program.
Just like in the <a href="fhog_object_detector_ex.cpp.html">fhog_object_detector_ex.cpp</a> example, we are going to train
a simple face detector based on the very small training dataset in the
examples/faces folder. As we will see, even with this small dataset the
MMOD method is able to make a working face detector. However, for real
applications you should train with more data for an even better result.
*/</font>
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>iostream<font color='#5555FF'>></font>
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>dlib<font color='#5555FF'>/</font>dnn.h<font color='#5555FF'>></font>
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>dlib<font color='#5555FF'>/</font>data_io.h<font color='#5555FF'>></font>
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>dlib<font color='#5555FF'>/</font>gui_widgets.h<font color='#5555FF'>></font>
<font color='#0000FF'>using</font> <font color='#0000FF'>namespace</font> std;
<font color='#0000FF'>using</font> <font color='#0000FF'>namespace</font> dlib;
<font color='#009900'>// The first thing we do is define our CNN. The CNN is going to be evaluated
</font><font color='#009900'>// convolutionally over an entire image pyramid. Think of it like a normal
</font><font color='#009900'>// sliding window classifier. This means you need to define a CNN that can look
</font><font color='#009900'>// at some part of an image and decide if it is an object of interest. In this
</font><font color='#009900'>// example I've defined a CNN with a receptive field of approximately 50x50
</font><font color='#009900'>// pixels. This is reasonable for face detection since you can clearly tell if
</font><font color='#009900'>// a 50x50 image contains a face. Other applications may benefit from CNNs with
</font><font color='#009900'>// different architectures.
</font><font color='#009900'>//
</font><font color='#009900'>// In this example our CNN begins with 3 downsampling layers. These layers will
</font><font color='#009900'>// reduce the size of the image by 8x and output a feature map with
</font><font color='#009900'>// 32 dimensions. Then we will pass that through 4 more convolutional layers to
</font><font color='#009900'>// get the final output of the network. The last layer has only 1 channel and
</font><font color='#009900'>// the values in that last channel are large when the network thinks it has
</font><font color='#009900'>// found an object at a particular location.
</font>
<font color='#009900'>// Let's begin the network definition by creating some network blocks.
</font>
<font color='#009900'>// A 5x5 conv layer that does 2x downsampling
</font><font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'><u>long</u></font> num_filters, <font color='#0000FF'>typename</font> SUBNET<font color='#5555FF'>></font> <font color='#0000FF'>using</font> con5d <font color='#5555FF'>=</font> con<font color='#5555FF'><</font>num_filters,<font color='#979000'>5</font>,<font color='#979000'>5</font>,<font color='#979000'>2</font>,<font color='#979000'>2</font>,SUBNET<font color='#5555FF'>></font>;
<font color='#009900'>// A 3x3 conv layer that doesn't do any downsampling
</font><font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'><u>long</u></font> num_filters, <font color='#0000FF'>typename</font> SUBNET<font color='#5555FF'>></font> <font color='#0000FF'>using</font> con3 <font color='#5555FF'>=</font> con<font color='#5555FF'><</font>num_filters,<font color='#979000'>3</font>,<font color='#979000'>3</font>,<font color='#979000'>1</font>,<font color='#979000'>1</font>,SUBNET<font color='#5555FF'>></font>;
<font color='#009900'>// Now we can define the 8x downsampling block in terms of conv5d blocks. We
</font><font color='#009900'>// also use relu and batch normalization in the standard way.
</font><font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'>typename</font> SUBNET<font color='#5555FF'>></font> <font color='#0000FF'>using</font> downsampler <font color='#5555FF'>=</font> relu<font color='#5555FF'><</font>bn_con<font color='#5555FF'><</font>con5d<font color='#5555FF'><</font><font color='#979000'>32</font>, relu<font color='#5555FF'><</font>bn_con<font color='#5555FF'><</font>con5d<font color='#5555FF'><</font><font color='#979000'>32</font>, relu<font color='#5555FF'><</font>bn_con<font color='#5555FF'><</font>con5d<font color='#5555FF'><</font><font color='#979000'>32</font>,SUBNET<font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font>;
<font color='#009900'>// The rest of the network will be 3x3 conv layers with batch normalization and
</font><font color='#009900'>// relu. So we define the 3x3 block we will use here.
</font><font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'>typename</font> SUBNET<font color='#5555FF'>></font> <font color='#0000FF'>using</font> rcon3 <font color='#5555FF'>=</font> relu<font color='#5555FF'><</font>bn_con<font color='#5555FF'><</font>con3<font color='#5555FF'><</font><font color='#979000'>32</font>,SUBNET<font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font>;
<font color='#009900'>// Finally, we define the entire network. The special input_rgb_image_pyramid
</font><font color='#009900'>// layer causes the network to operate over a spatial pyramid, making the detector
</font><font color='#009900'>// scale invariant.
</font><font color='#0000FF'>using</font> net_type <font color='#5555FF'>=</font> loss_mmod<font color='#5555FF'><</font>con<font color='#5555FF'><</font><font color='#979000'>1</font>,<font color='#979000'>6</font>,<font color='#979000'>6</font>,<font color='#979000'>1</font>,<font color='#979000'>1</font>,rcon3<font color='#5555FF'><</font>rcon3<font color='#5555FF'><</font>rcon3<font color='#5555FF'><</font>downsampler<font color='#5555FF'><</font>input_rgb_image_pyramid<font color='#5555FF'><</font>pyramid_down<font color='#5555FF'><</font><font color='#979000'>6</font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font>;
<font color='#009900'>// ----------------------------------------------------------------------------------------
</font>
<font color='#0000FF'><u>int</u></font> <b><a name='main'></a>main</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>int</u></font> argc, <font color='#0000FF'><u>char</u></font><font color='#5555FF'>*</font><font color='#5555FF'>*</font> argv<font face='Lucida Console'>)</font> <font color='#0000FF'>try</font>
<b>{</b>
<font color='#009900'>// In this example we are going to train a face detector based on the
</font> <font color='#009900'>// small faces dataset in the examples/faces directory. So the first
</font> <font color='#009900'>// thing we do is load that dataset. This means you need to supply the
</font> <font color='#009900'>// path to this faces folder as a command line argument so we will know
</font> <font color='#009900'>// where it is.
</font> <font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>argc <font color='#5555FF'>!</font><font color='#5555FF'>=</font> <font color='#979000'>2</font><font face='Lucida Console'>)</font>
<b>{</b>
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>Give the path to the examples/faces directory as the argument to this</font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>program. For example, if you are in the examples folder then execute </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>this program by running: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'> ./dnn_mmod_ex faces</font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#0000FF'>return</font> <font color='#979000'>0</font>;
<b>}</b>
<font color='#0000FF'>const</font> std::string faces_directory <font color='#5555FF'>=</font> argv[<font color='#979000'>1</font>];
<font color='#009900'>// The faces directory contains a training dataset and a separate
</font> <font color='#009900'>// testing dataset. The training data consists of 4 images, each
</font> <font color='#009900'>// annotated with rectangles that bound each human face. The idea is
</font> <font color='#009900'>// to use this training data to learn to identify human faces in new
</font> <font color='#009900'>// images.
</font> <font color='#009900'>//
</font> <font color='#009900'>// Once you have trained an object detector it is always important to
</font> <font color='#009900'>// test it on data it wasn't trained on. Therefore, we will also load
</font> <font color='#009900'>// a separate testing set of 5 images. Once we have a face detector
</font> <font color='#009900'>// created from the training data we will see how well it works by
</font> <font color='#009900'>// running it on the testing images.
</font> <font color='#009900'>//
</font> <font color='#009900'>// So here we create the variables that will hold our dataset.
</font> <font color='#009900'>// images_train will hold the 4 training images and face_boxes_train
</font> <font color='#009900'>// holds the locations of the faces in the training images. So for
</font> <font color='#009900'>// example, the image images_train[0] has the faces given by the
</font> <font color='#009900'>// rectangles in face_boxes_train[0].
</font> std::vector<font color='#5555FF'><</font>matrix<font color='#5555FF'><</font>rgb_pixel<font color='#5555FF'>></font><font color='#5555FF'>></font> images_train, images_test;
std::vector<font color='#5555FF'><</font>std::vector<font color='#5555FF'><</font>mmod_rect<font color='#5555FF'>></font><font color='#5555FF'>></font> face_boxes_train, face_boxes_test;
<font color='#009900'>// Now we load the data. These XML files list the images in each dataset
</font> <font color='#009900'>// and also contain the positions of the face boxes. Obviously you can use
</font> <font color='#009900'>// any kind of input format you like so long as you store the data into
</font> <font color='#009900'>// images_train and face_boxes_train. But for convenience dlib comes with
</font> <font color='#009900'>// tools for creating and loading XML image datasets. Here you see how to
</font> <font color='#009900'>// load the data. To create the XML files you can use the imglab tool which
</font> <font color='#009900'>// can be found in the tools/imglab folder. It is a simple graphical tool
</font> <font color='#009900'>// for labeling objects in images with boxes. To see how to use it read the
</font> <font color='#009900'>// tools/imglab/README.txt file.
</font> <font color='#BB00BB'>load_image_dataset</font><font face='Lucida Console'>(</font>images_train, face_boxes_train, faces_directory<font color='#5555FF'>+</font>"<font color='#CC0000'>/training.xml</font>"<font face='Lucida Console'>)</font>;
<font color='#BB00BB'>load_image_dataset</font><font face='Lucida Console'>(</font>images_test, face_boxes_test, faces_directory<font color='#5555FF'>+</font>"<font color='#CC0000'>/testing.xml</font>"<font face='Lucida Console'>)</font>;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>num training images: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> images_train.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>num testing images: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> images_test.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// The MMOD algorithm has some options you can set to control its behavior. However,
</font> <font color='#009900'>// you can also call the constructor with your training annotations and a "target
</font> <font color='#009900'>// object size" and it will automatically configure itself in a reasonable way for your
</font> <font color='#009900'>// problem. Here we are saying that faces are still recognizably faces when they are
</font> <font color='#009900'>// 40x40 pixels in size. You should generally pick the smallest size where this is
</font> <font color='#009900'>// true. Based on this information the mmod_options constructor will automatically
</font> <font color='#009900'>// pick a good sliding window width and height. It will also automatically set the
</font> <font color='#009900'>// non-max-suppression parameters to something reasonable. For further details see the
</font> <font color='#009900'>// mmod_options documentation.
</font> mmod_options <font color='#BB00BB'>options</font><font face='Lucida Console'>(</font>face_boxes_train, <font color='#979000'>40</font>,<font color='#979000'>40</font><font face='Lucida Console'>)</font>;
<font color='#009900'>// The detector will automatically decide to use multiple sliding windows if needed.
</font> <font color='#009900'>// For the face data, only one is needed however.
</font> cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>num detector windows: </font>"<font color='#5555FF'><</font><font color='#5555FF'><</font> options.detector_windows.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font> w : options.detector_windows<font face='Lucida Console'>)</font>
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>detector window width by height: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> w.width <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'> x </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> w.height <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>overlap NMS IOU thresh: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> options.overlaps_nms.<font color='#BB00BB'>get_iou_thresh</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>overlap NMS percent covered thresh: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> options.overlaps_nms.<font color='#BB00BB'>get_percent_covered_thresh</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// Now we are ready to create our network and trainer.
</font> net_type <font color='#BB00BB'>net</font><font face='Lucida Console'>(</font>options<font face='Lucida Console'>)</font>;
<font color='#009900'>// The MMOD loss requires that the number of filters in the final network layer equal
</font> <font color='#009900'>// options.detector_windows.size(). So we set that here as well.
</font> net.<font color='#BB00BB'>subnet</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>.<font color='#BB00BB'>layer_details</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>.<font color='#BB00BB'>set_num_filters</font><font face='Lucida Console'>(</font>options.detector_windows.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>;
dnn_trainer<font color='#5555FF'><</font>net_type<font color='#5555FF'>></font> <font color='#BB00BB'>trainer</font><font face='Lucida Console'>(</font>net<font face='Lucida Console'>)</font>;
trainer.<font color='#BB00BB'>set_learning_rate</font><font face='Lucida Console'>(</font><font color='#979000'>0.1</font><font face='Lucida Console'>)</font>;
trainer.<font color='#BB00BB'>be_verbose</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>;
trainer.<font color='#BB00BB'>set_synchronization_file</font><font face='Lucida Console'>(</font>"<font color='#CC0000'>mmod_sync</font>", std::chrono::<font color='#BB00BB'>minutes</font><font face='Lucida Console'>(</font><font color='#979000'>5</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>;
trainer.<font color='#BB00BB'>set_iterations_without_progress_threshold</font><font face='Lucida Console'>(</font><font color='#979000'>300</font><font face='Lucida Console'>)</font>;
<font color='#009900'>// Now let's train the network. We are going to use mini-batches of 150
</font> <font color='#009900'>// images. The images are random crops from our training set (see
</font> <font color='#009900'>// <a href="random_cropper_ex.cpp.html">random_cropper_ex.cpp</a> for a discussion of the random_cropper).
</font> std::vector<font color='#5555FF'><</font>matrix<font color='#5555FF'><</font>rgb_pixel<font color='#5555FF'>></font><font color='#5555FF'>></font> mini_batch_samples;
std::vector<font color='#5555FF'><</font>std::vector<font color='#5555FF'><</font>mmod_rect<font color='#5555FF'>></font><font color='#5555FF'>></font> mini_batch_labels;
random_cropper cropper;
cropper.<font color='#BB00BB'>set_chip_dims</font><font face='Lucida Console'>(</font><font color='#979000'>200</font>, <font color='#979000'>200</font><font face='Lucida Console'>)</font>;
<font color='#009900'>// Usually you want to give the cropper whatever min sizes you passed to the
</font> <font color='#009900'>// mmod_options constructor, which is what we do here.
</font> cropper.<font color='#BB00BB'>set_min_object_size</font><font face='Lucida Console'>(</font><font color='#979000'>40</font>,<font color='#979000'>40</font><font face='Lucida Console'>)</font>;
dlib::rand rnd;
<font color='#009900'>// Run the trainer until the learning rate gets small. This will probably take several
</font> <font color='#009900'>// hours.
</font> <font color='#0000FF'>while</font><font face='Lucida Console'>(</font>trainer.<font color='#BB00BB'>get_learning_rate</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'>></font><font color='#5555FF'>=</font> <font color='#979000'>1e</font><font color='#5555FF'>-</font><font color='#979000'>4</font><font face='Lucida Console'>)</font>
<b>{</b>
<font color='#BB00BB'>cropper</font><font face='Lucida Console'>(</font><font color='#979000'>150</font>, images_train, face_boxes_train, mini_batch_samples, mini_batch_labels<font face='Lucida Console'>)</font>;
<font color='#009900'>// We can also randomly jitter the colors and that often helps a detector
</font> <font color='#009900'>// generalize better to new images.
</font> <font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> img : mini_batch_samples<font face='Lucida Console'>)</font>
<font color='#BB00BB'>disturb_colors</font><font face='Lucida Console'>(</font>img, rnd<font face='Lucida Console'>)</font>;
trainer.<font color='#BB00BB'>train_one_step</font><font face='Lucida Console'>(</font>mini_batch_samples, mini_batch_labels<font face='Lucida Console'>)</font>;
<b>}</b>
<font color='#009900'>// wait for training threads to stop
</font> trainer.<font color='#BB00BB'>get_net</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>;
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>done training</font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// Save the network to disk
</font> net.<font color='#BB00BB'>clean</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>;
<font color='#BB00BB'>serialize</font><font face='Lucida Console'>(</font>"<font color='#CC0000'>mmod_network.dat</font>"<font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> net;
<font color='#009900'>// Now that we have a face detector we can test it. The first statement tests it
</font> <font color='#009900'>// on the training data. It will print the precision, recall, and then average precision.
</font> <font color='#009900'>// This statement should indicate that the network works perfectly on the
</font> <font color='#009900'>// training data.
</font> cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>training results: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> <font color='#BB00BB'>test_object_detection_function</font><font face='Lucida Console'>(</font>net, images_train, face_boxes_train<font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// However, to get an idea if it really worked without overfitting we need to run
</font> <font color='#009900'>// it on images it wasn't trained on. The next line does this. Happily,
</font> <font color='#009900'>// this statement indicates that the detector finds most of the faces in the
</font> <font color='#009900'>// testing data.
</font> cout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>testing results: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> <font color='#BB00BB'>test_object_detection_function</font><font face='Lucida Console'>(</font>net, images_test, face_boxes_test<font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// If you are running many experiments, it's also useful to log the settings used
</font> <font color='#009900'>// during the training experiment. This statement will print the settings we used to
</font> <font color='#009900'>// the screen.
</font> cout <font color='#5555FF'><</font><font color='#5555FF'><</font> trainer <font color='#5555FF'><</font><font color='#5555FF'><</font> cropper <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<font color='#009900'>// Now lets run the detector on the testing images and look at the outputs.
</font> image_window win;
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> img : images_test<font face='Lucida Console'>)</font>
<b>{</b>
<font color='#BB00BB'>pyramid_up</font><font face='Lucida Console'>(</font>img<font face='Lucida Console'>)</font>;
<font color='#0000FF'>auto</font> dets <font color='#5555FF'>=</font> <font color='#BB00BB'>net</font><font face='Lucida Console'>(</font>img<font face='Lucida Console'>)</font>;
win.<font color='#BB00BB'>clear_overlay</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>;
win.<font color='#BB00BB'>set_image</font><font face='Lucida Console'>(</font>img<font face='Lucida Console'>)</font>;
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> d : dets<font face='Lucida Console'>)</font>
win.<font color='#BB00BB'>add_overlay</font><font face='Lucida Console'>(</font>d<font face='Lucida Console'>)</font>;
cin.<font color='#BB00BB'>get</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>;
<b>}</b>
<font color='#0000FF'>return</font> <font color='#979000'>0</font>;
<font color='#009900'>// Now that you finished this example, you should read <a href="dnn_mmod_train_find_cars_ex.cpp.html">dnn_mmod_train_find_cars_ex.cpp</a>,
</font> <font color='#009900'>// which is a more advanced example. It discusses many issues surrounding properly
</font> <font color='#009900'>// setting the MMOD parameters and creating a good training dataset.
</font>
<b>}</b>
<font color='#0000FF'>catch</font><font face='Lucida Console'>(</font>std::exception<font color='#5555FF'>&</font> e<font face='Lucida Console'>)</font>
<b>{</b>
cout <font color='#5555FF'><</font><font color='#5555FF'><</font> e.<font color='#BB00BB'>what</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font><font color='#5555FF'><</font> endl;
<b>}</b>
</pre></body></html>