dnn_mmod_train_find_cars_ex.cpp

// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
/*
    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 fhog_object_detector_ex.cpp) 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 dnn_introduction_ex.cpp and dnn_introduction2_ex.cpp
    before reading this example program.  You should also read the introductory DNN+MMOD
    example dnn_mmod_ex.cpp as well before proceeding.
    

    This example is essentially a more complex version of dnn_mmod_ex.cpp.  In it we train
    a detector that finds the rear ends of motor vehicles.  I will also discuss some
    aspects of data preparation useful when training this kind of detector.  
    
*/


#include <iostream>
#include <dlib/dnn.h>
#include <dlib/data_io.h>

using namespace std;
using namespace dlib;



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<bn_con<con5d<32, relu<bn_con<con5d<32, relu<bn_con<con5d<16,SUBNET>>>>>>>>>;
template <typename SUBNET> using rcon5  = relu<bn_con<con5<55,SUBNET>>>;
using net_type = loss_mmod<con<1,9,9,1,1,rcon5<rcon5<rcon5<downsampler<input_rgb_image_pyramid<pyramid_down<6>>>>>>>>;


// ----------------------------------------------------------------------------------------

int ignore_overlapped_boxes(
    std::vector<mmod_rect>& boxes,
    const test_box_overlap& overlaps
)
/*!
    ensures
        - Whenever two rectangles in boxes overlap, according to overlaps(), we set the
          smallest box to ignore.
        - returns the number of newly ignored boxes.
!*/
{
    int num_ignored = 0;
    for (size_t i = 0; i < boxes.size(); ++i)
    {
        if (boxes[i].ignore)
            continue;
        for (size_t j = i+1; j < boxes.size(); ++j)
        {
            if (boxes[j].ignore)
                continue;
            if (overlaps(boxes[i], boxes[j]))
            {
                ++num_ignored;
                if(boxes[i].rect.area() < boxes[j].rect.area())
                    boxes[i].ignore = true;
                else
                    boxes[j].ignore = true;
            }
        }
    }
    return num_ignored;
}

// ----------------------------------------------------------------------------------------

int main(int argc, char** argv) try
{
    if (argc != 2)
    {
        cout << "Give the path to a folder containing training.xml and testing.xml files." << endl;
        cout << "This example program is specifically designed to run on the dlib vehicle " << endl;
        cout << "detection dataset, which is available at this URL: " << endl;
        cout << "   http://dlib.net/files/data/dlib_rear_end_vehicles_v1.tar" << endl;
        cout << endl;
        cout << "So download that dataset, extract it somewhere, and then run this program" << endl;
        cout << "with the dlib_rear_end_vehicles folder as an argument.  E.g. if you extract" << endl;
        cout << "the dataset to the current folder then you should run this example program" << endl;
        cout << "by typing: " << endl;
        cout << "   ./dnn_mmod_train_find_cars_ex dlib_rear_end_vehicles" << endl;
        cout << endl;
        cout << "It takes about a day to finish if run on a high end GPU like a 1080ti." << endl;
        cout << endl;
        return 0;
    }
    const std::string data_directory = argv[1];


    std::vector<matrix<rgb_pixel>> images_train, images_test;
    std::vector<std::vector<mmod_rect>> boxes_train, boxes_test;
    load_image_dataset(images_train, boxes_train, data_directory+"/training.xml");
    load_image_dataset(images_test,  boxes_test,  data_directory+"/testing.xml");

    // When I was creating the dlib vehicle detection dataset I had to label all the cars
    // in each image.  MMOD requires all cars to be labeled, since any unlabeled part of an
    // image is implicitly assumed to be not a car, and the algorithm will use it as
    // negative training data.  So every car must be labeled, either with a normal
    // rectangle or an "ignore" rectangle that tells MMOD to simply ignore it (i.e. neither
    // treat it as a thing to detect nor as negative training data).  
    // 
    // In our present case, many images contain very tiny cars in the distance, ones that
    // are essentially just dark smudges.  It's not reasonable to expect the CNN
    // architecture we defined to detect such vehicles.  However, I erred on the side of
    // having more complete annotations when creating the dataset.  So when I labeled these
    // images I labeled many of these really difficult cases as vehicles to detect.   
    //
    // So the first thing we are going to do is clean up our dataset a little bit.  In
    // particular, we are going to mark boxes smaller than 35*35 pixels as ignore since
    // only really small and blurry cars appear at those sizes.  We will also mark boxes
    // that are heavily overlapped by another box as ignore.  We do this because we want to
    // allow for stronger non-maximum suppression logic in the learned detector, since that
    // will help make it easier to learn a good detector. 
    // 
    // To explain this non-max suppression idea further it's important to understand how
    // the detector works.  Essentially, sliding window detectors scan all image locations
    // and ask "is there a care here?".  If there really is a car in a specific location in
    // an image then usually many slightly different sliding window locations will produce
    // high detection scores, indicating that there is a car at those locations.  If we
    // just stopped there then each car would produce multiple detections.  But that isn't
    // what we want.  We want each car to produce just one detection.  So it's common for
    // detectors to include "non-maximum suppression" logic which simply takes the
    // strongest detection and then deletes all detections "close to" the strongest.  This
    // is a simple post-processing step that can eliminate duplicate detections.  However,
    // we have to define what "close to" means.  We can do this by looking at your training
    // data and checking how close the closest target boxes are to each other, and then
    // picking a "close to" measure that doesn't suppress those target boxes but is
    // otherwise as tight as possible.  This is exactly what the mmod_options object does
    // by default.
    //
    // Importantly, this means that if your training dataset contains an image with two
    // target boxes that really overlap a whole lot, then the non-maximum suppression
    // "close to" measure will be configured to allow detections to really overlap a whole
    // lot.  On the other hand, if your dataset didn't contain any overlapped boxes at all,
    // then the non-max suppression logic would be configured to filter out any boxes that
    // overlapped at all, and thus would be performing a much stronger non-max suppression.  
    //
    // Why does this matter?  Well, remember that we want to avoid duplicate detections.
    // If non-max suppression just kills everything in a really wide area around a car then
    // the CNN doesn't really need to learn anything about avoiding duplicate detections.
    // However, if non-max suppression only suppresses a tiny area around each detection
    // then the CNN will need to learn to output small detection scores for those areas of
    // the image not suppressed.  The smaller the non-max suppression region the more the
    // CNN has to learn and the more difficult the learning problem will become.  This is
    // why we remove highly overlapped objects from the training dataset.  That is, we do
    // it so the non-max suppression logic will be able to be reasonably effective.  Here
    // we are ensuring that any boxes that are entirely contained by another are
    // suppressed.  We also ensure that boxes with an intersection over union of 0.5 or
    // greater are suppressed.  This will improve the resulting detector since it will be
    // able to use more aggressive non-max suppression settings.

    int num_overlapped_ignored_test = 0;
    for (auto& v : boxes_test)
        num_overlapped_ignored_test += ignore_overlapped_boxes(v, test_box_overlap(0.50, 0.95));

    int num_overlapped_ignored = 0;
    int num_additional_ignored = 0;
    for (auto& v : boxes_train)
    {
        num_overlapped_ignored += ignore_overlapped_boxes(v, test_box_overlap(0.50, 0.95));
        for (auto& bb : v)
        {
            if (bb.rect.width() < 35 && bb.rect.height() < 35)
            {
                if (!bb.ignore)
                {
                    bb.ignore = true;
                    ++num_additional_ignored;
                }
            }

            // The dlib vehicle detection dataset doesn't contain any detections with
            // really extreme aspect ratios.  However, some datasets do, often because of
            // bad labeling.  So it's a good idea to check for that and either eliminate
            // those boxes or set them to ignore.  Although, this depends on your
            // application.  
            // 
            // For instance, if your dataset has boxes with an aspect ratio
            // of 10 then you should think about what that means for the network
            // architecture.  Does the receptive field even cover the entirety of the box
            // in those cases?  Do you care about these boxes?  Are they labeling errors?
            // I find that many people will download some dataset from the internet and
            // just take it as given.  They run it through some training algorithm and take
            // the dataset as unchallengeable truth.  But many datasets are full of
            // labeling errors.  There are also a lot of datasets that aren't full of
            // errors, but are annotated in a sloppy and inconsistent way.  Fixing those
            // errors and inconsistencies can often greatly improve models trained from
            // such data.  It's almost always worth the time to try and improve your
            // training dataset.   
            //
            // In any case, my point is that there are other types of dataset cleaning you
            // could put here.  What exactly you need depends on your application.  But you
            // should carefully consider it and not take your dataset as a given.  The work
            // of creating a good detector is largely about creating a high quality
            // training dataset.  
        }
    }

    // When modifying a dataset like this, it's a really good idea to print a log of how
    // many boxes you ignored.  It's easy to accidentally ignore a huge block of data, so
    // you should always look and see that things are doing what you expect.
    cout << "num_overlapped_ignored: "<< num_overlapped_ignored << endl;
    cout << "num_additional_ignored: "<< num_additional_ignored << endl;
    cout << "num_overlapped_ignored_test: "<< num_overlapped_ignored_test << endl;


    cout << "num training images: " << images_train.size() << endl;
    cout << "num testing images: " << images_test.size() << endl;


    // Our vehicle detection dataset has basically 3 different types of boxes.  Square
    // boxes, tall and skinny boxes (e.g. semi trucks), and short and wide boxes (e.g.
    // sedans).  Here we are telling the MMOD algorithm that a vehicle is recognizable as
    // long as the longest box side is at least 70 pixels long and the shortest box side is
    // at least 30 pixels long.  mmod_options will use these parameters to decide how large
    // each of the sliding windows needs to be so as to be able to detect all the vehicles.
    // Since our dataset has basically these 3 different aspect ratios, it will decide to
    // use 3 different sliding windows.  This means the final con layer in the network will
    // have 3 filters, one for each of these aspect ratios. 
    //
    // Another thing to consider when setting the sliding window size is the "stride" of
    // your network.  The network we defined above downsamples the image by a factor of 8x
    // in the first few layers.  So when the sliding windows are scanning the image, they
    // are stepping over it with a stride of 8 pixels.  If you set the sliding window size
    // too small then the stride will become an issue.  For instance, if you set the
    // sliding window size to 4 pixels, then it means a 4x4 window will be moved by 8
    // pixels at a time when scanning. This is obviously a problem since 75% of the image
    // won't even be visited by the sliding window.  So you need to set the window size to
    // be big enough relative to the stride of your network.  In our case, the windows are
    // at least 30 pixels in length, so being moved by 8 pixel steps is fine. 
    mmod_options options(boxes_train, 70, 30);


    // This setting is very important and dataset specific.  The vehicle detection dataset
    // contains boxes that are marked as "ignore", as we discussed above.  Some of them are
    // ignored because we set ignore to true in the above code.  However, the xml files
    // also contained a lot of ignore boxes.  Some of them are large boxes that encompass
    // large parts of an image and the intention is to have everything inside those boxes
    // be ignored.  Therefore, we need to tell the MMOD algorithm to do that, which we do
    // by setting options.overlaps_ignore appropriately.  
    // 
    // But first, we need to understand exactly what this option does.  The MMOD loss
    // is essentially counting the number of false alarms + missed detections produced by
    // the detector for each image.  During training, the code is running the detector on
    // each image in a mini-batch and looking at its output and counting the number of
    // mistakes.  The optimizer tries to find parameters settings that minimize the number
    // of detector mistakes.
    // 
    // This overlaps_ignore option allows you to tell the loss that some outputs from the
    // detector should be totally ignored, as if they never happened.  In particular, if a
    // detection overlaps a box in the training data with ignore==true then that detection
    // is ignored.  This overlap is determined by calling
    // options.overlaps_ignore(the_detection, the_ignored_training_box).  If it returns
    // true then that detection is ignored.
    // 
    // You should read the documentation for test_box_overlap, the class type for
    // overlaps_ignore for full details.  However, the gist is that the default behavior is
    // to only consider boxes as overlapping if their intersection over union is > 0.5.
    // However, the dlib vehicle detection dataset contains large boxes that are meant to
    // mask out large areas of an image.  So intersection over union isn't an appropriate
    // way to measure "overlaps with box" in this case.  We want any box that is contained
    // inside one of these big regions to be ignored, even if the detection box is really
    // small.  So we set overlaps_ignore to behave that way with this line.
    options.overlaps_ignore = test_box_overlap(0.5, 0.95);

    net_type net(options);

    // The final layer of the network must be a con layer that contains 
    // options.detector_windows.size() filters.  This is because these final filters are
    // what perform the final "sliding window" detection in the network.  For the dlib
    // vehicle dataset, there will be 3 sliding window detectors, so we will be setting
    // num_filters to 3 here.
    net.subnet().layer_details().set_num_filters(options.detector_windows.size());


    dnn_trainer<net_type> trainer(net,sgd(0.0001,0.9));
    trainer.set_learning_rate(0.1);
    trainer.be_verbose();


    // While training, we are going to use early stopping.  That is, we will be checking
    // how good the detector is performing on our test data and when it stops getting
    // better on the test data we will drop the learning rate.  We will keep doing that
    // until the learning rate is less than 1e-4.   These two settings tell the trainer to
    // do that.  Essentially, we are setting the first argument to infinity, and only the
    // test iterations without progress threshold will matter.  In particular, it says that
    // once we observe 1000 testing mini-batches where the test loss clearly isn't
    // decreasing we will lower the learning rate.
    trainer.set_iterations_without_progress_threshold(50000);
    trainer.set_test_iterations_without_progress_threshold(1000);

    const string sync_filename = "mmod_cars_sync";
    trainer.set_synchronization_file(sync_filename, std::chrono::minutes(5));




    std::vector<matrix<rgb_pixel>> mini_batch_samples;
    std::vector<std::vector<mmod_rect>> mini_batch_labels; 
    random_cropper cropper;
    cropper.set_seed(time(0));
    cropper.set_chip_dims(350, 350);
    cropper.set_min_object_size(0.20); 
    cropper.set_max_rotation_degrees(2);
    dlib::rand rnd;

    // Log the training parameters to the console
    cout << trainer << cropper << endl;

    int cnt = 1;
    // Run the trainer until the learning rate gets small.  
    while(trainer.get_learning_rate() >= 1e-4)
    {
        // Every 30 mini-batches we do a testing mini-batch.  
        if (cnt%30 != 0 || images_test.size() == 0)
        {
            cropper(87, images_train, boxes_train, mini_batch_samples, mini_batch_labels);
            // We can also randomly jitter the colors and that often helps a detector
            // generalize better to new images.
            for (auto&& img : mini_batch_samples)
                disturb_colors(img, rnd);

            // It's a good idea to, at least once, put code here that displays the images
            // and boxes the random cropper is generating.  You should look at them and
            // think about if the output makes sense for your problem.  Most of the time
            // it will be fine, but sometimes you will realize that the pattern of cropping
            // isn't really appropriate for your problem and you will need to make some
            // change to how the mini-batches are being generated.  Maybe you will tweak
            // some of the cropper's settings, or write your own entirely separate code to
            // create mini-batches.  But either way, if you don't look you will never know.
            // An easy way to do this is to create a dlib::image_window to display the
            // images and boxes.

            trainer.train_one_step(mini_batch_samples, mini_batch_labels);
        }
        else
        {
            cropper(87, images_test, boxes_test, mini_batch_samples, mini_batch_labels);
            // We can also randomly jitter the colors and that often helps a detector
            // generalize better to new images.
            for (auto&& img : mini_batch_samples)
                disturb_colors(img, rnd);

            trainer.test_one_step(mini_batch_samples, mini_batch_labels);
        }
        ++cnt;
    }
    // wait for training threads to stop
    trainer.get_net();
    cout << "done training" << endl;

    // Save the network to disk
    net.clean();
    serialize("mmod_rear_end_vehicle_detector.dat") << net;


    // It's a really good idea to print the training parameters.  This is because you will
    // invariably be running multiple rounds of training and should be logging the output
    // to a file.  This print statement will include many of the training parameters in
    // your log.
    cout << trainer << cropper << endl;

    cout << "\nsync_filename: " << sync_filename << endl;
    cout << "num training images: "<< images_train.size() << endl;
    cout << "training results: " << test_object_detection_function(net, images_train, boxes_train, test_box_overlap(), 0, options.overlaps_ignore);
    // Upsampling the data will allow the detector to find smaller cars.  Recall that 
    // we configured it to use a sliding window nominally 70 pixels in size.  So upsampling
    // here will let it find things nominally 35 pixels in size.  Although we include a
    // limit of 1800*1800 here which means "don't upsample an image if it's already larger
    // than 1800*1800".  We do this so we don't run out of RAM, which is a concern because
    // some of the images in the dlib vehicle dataset are really high resolution.
    upsample_image_dataset<pyramid_down<2>>(images_train, boxes_train, 1800*1800);
    cout << "training upsampled results: " << test_object_detection_function(net, images_train, boxes_train, test_box_overlap(), 0, options.overlaps_ignore);


    cout << "num testing images: "<< images_test.size() << endl;
    cout << "testing results: " << test_object_detection_function(net, images_test, boxes_test, test_box_overlap(), 0, options.overlaps_ignore);
    upsample_image_dataset<pyramid_down<2>>(images_test, boxes_test, 1800*1800);
    cout << "testing upsampled results: " << test_object_detection_function(net, images_test, boxes_test, test_box_overlap(), 0, options.overlaps_ignore);

    /*
        This program takes many hours to execute on a high end GPU.  It took about a day to
        train on a NVIDIA 1080ti.  The resulting model file is available at
            http://dlib.net/files/mmod_rear_end_vehicle_detector.dat.bz2
        It should be noted that this file on dlib.net has a dlib::shape_predictor appended
        onto the end of it (see dnn_mmod_find_cars_ex.cpp for an example of its use).  This
        explains why the model file on dlib.net is larger than the
        mmod_rear_end_vehicle_detector.dat output by this program.

        You can see some videos of this vehicle detector running on YouTube:
            https://www.youtube.com/watch?v=4B3bzmxMAZU
            https://www.youtube.com/watch?v=bP2SUo5vSlc

        Also, the training and testing accuracies were:
            num training images: 2217
            training results: 0.990738 0.736431 0.736073 
            training upsampled results: 0.986837 0.937694 0.936912 
            num testing images: 135
            testing results: 0.988827 0.471372 0.470806 
            testing upsampled results: 0.987879 0.651132 0.650399 
    */

    return 0;

}
catch(std::exception& e)
{
    cout << e.what() << endl;
}