How to identify stranger in dlib`s one vs one classifier - c++

Question

I`m using a one_vs_one_trainer and one_vs_one_decision_function for classify 128D face descriptors, and i want to detect unknown face.
I`m detecting faces using OpenCV and my wrapper, then i followed the guide and computed the 128D face descriptors, that i stored in files. Next, i trained one_vs_one classifier following this tutorial. All works perfectly, but when i try to classify unknown face it returns some label.
I used code from guides, but if you want to look at my code - it is here
Is there a better way to identify faces? Maybe, its simpler to use OpenCV`s methods, or other from Dlib?

Answer 1

Thanks for Davis!
Here is forum thread on SourceForge.
The answer is:
Use a bunch of binary classifiers rather than one vs one. If all the binary
classifiers say they don't match then you know the person doesn't match any
of them.
And i implemented this as follows:
#include <iostream>
#include <ctime>
#include <vector>
#include <dlib/svm.h>
using namespace std;
using namespace dlib;
int main() {
typedef matrix<double, 128, 1> sample_type;
typedef histogram_intersection_kernel<sample_type> kernel_type;
typedef svm_c_trainer<kernel_type> trainer_type;
typedef decision_function<kernel_type> classifier_type;
std::vector<sample_type> samples;
std::vector<double> labels;
sample_type sample;
// Samples ->
sample = -0.104075,0.0353173,...,0.114782,-0.0360935;
samples.emplace_back(sample);
labels.emplace_back(0);
sample = -0.0842,-0.0103397,...,0.0938285,0.010045;
samples.emplace_back(sample);
labels.emplace_back(0);
sample = -0.0978358,0.0709425,...,0.052436,-0.0582029;
samples.emplace_back(sample);
labels.emplace_back(0);
sample = -0.126522,0.0319873,...,0.12045,-0.0277105;
samples.emplace_back(sample);
labels.emplace_back(0);
sample = -0.10335,-0.0261625,...,0.0600661,0.00703168,-8.67462e-05,-0.0598214,-0.104442,-0.046698,0.0553857,-0.0880691,0.0482511,0.0331484;
samples.emplace_back(sample);
labels.emplace_back(0);
sample = -0.0747794,0.0599716,...,-0.0440207,-6.45183e-05;
samples.emplace_back(sample);
labels.emplace_back(1);
sample = -0.0280804,0.0900723,...,-0.0267513,0.00824318;
samples.emplace_back(sample);
labels.emplace_back(1);
sample = -0.0721213,0.00700722,...,-0.0128318,0.100784;
samples.emplace_back(sample);
labels.emplace_back(1);
sample = -0.122747,0.0737782,0.0375799,...,0.0168201,-0.0246723;
samples.emplace_back(sample);
labels.emplace_back(1);
sample = -0.0218071,0.118063,...,-0.0735178,0.04046;
samples.emplace_back(sample);
labels.emplace_back(1);
sample = -0.0680787,0.0490121,-0.0228516,...,-0.0366242,0.0287891;
samples.emplace_back(sample);
labels.emplace_back(2);
sample = 0.00152394,0.107174,...,-0.0479925,0.0182667;
samples.emplace_back(sample);
labels.emplace_back(2);
sample = -0.0334521,0.165314,...,-0.0385227,-0.0215499;
samples.emplace_back(sample);
labels.emplace_back(2);
sample = 0.0276394,0.106774,...,-0.0496831,-0.020857;
samples.emplace_back(sample);
labels.emplace_back(2);
// <- Samples
// Unique labels ->
std::vector<double> total_labels;
for(double &label : labels) {
if(find(total_labels.begin(), total_labels.end(), label) == total_labels.end())
total_labels.emplace_back(label);
}
// <- Unique labels
// Init trainers ->
std::vector<trainer_type> trainers;
int num_trainers = total_labels.size() * (total_labels.size() - 1) / 2;
cout << "Number of trainers is " << num_trainers << endl;
for(int i = 0; i < num_trainers; i++) {
trainers.emplace_back(trainer_type());
trainers[i].set_kernel(kernel_type());
trainers[i].set_c(10);
}
// <- Init trainers
// Init classifiers ->
std::vector<pair<double, double>> classifiersLabels;
std::vector<classifier_type> classifiers;
int label1 = 0, label2 = 1;
for(trainer_type &trainer : trainers) {
std::vector<sample_type> samples4pair;
std::vector<double> labels4pair;
for(int i = 0; i < samples.size(); i++) {
if(labels[i] == total_labels[label1]) {
samples4pair.emplace_back(samples[i]);
labels4pair.emplace_back(-1);
}
if(labels[i] == total_labels[label2]) {
samples4pair.emplace_back(samples[i]);
labels4pair.emplace_back(+1);
}
}
classifiers.emplace_back(trainer.train(samples4pair, labels4pair));
classifiersLabels.emplace_back(make_pair(total_labels[label1],
total_labels[label2]));
label2++;
if(label2 == total_labels.size()) {
label1++;
label2 = label1 + 1;
}
}
// <- Init classifiers
double threshold = 0.3;
auto classify = [&](){
std::map<double, int> votes;
for(int i = 0; i < classifiers.size(); i++) {
cout << "Classifier #" << i << ":" << endl;
double prediction = classifiers[i](sample);
cout << prediction << ": ";
if(abs(prediction) < threshold) {
cout << "-1" << endl;
} else if (prediction < 0) {
votes[classifiersLabels[i].first]++;
cout << classifiersLabels[i].first << endl;
} else {
votes[classifiersLabels[i].second]++;
cout << classifiersLabels[i].second << endl;
}
}
cout << "Votes: " << endl;
for(auto &vote : votes) {
cout << vote.first << ": " << vote.second << endl;
}
auto max = std::max_element(votes.begin(), votes.end(),
[](const pair<double, int>& p1, const pair<double, int>& p2) {
return p1.second < p2.second; });
double label = votes.empty() ? -1 : max->first;
cout << "Label is " << label << endl;
};
// Test ->
cout << endl;
sample = -0.0971093, ..., 0.123482, -0.0399552;
cout << "True: 0 - " << endl;
classify();
cout << endl;
sample = -0.0548414, ..., 0.0277335, 0.0460183;
cout << "True: 1 - " << endl;
classify();
cout << endl;
sample = -0.0456186,0.0617834,...,-0.0387607,0.0366309;
cout << "True: 1 - " << endl;
classify();
cout << endl;
sample = -0.0500396, 0.0947202, ..., -0.0540899, 0.0206803;
cout << "True: 2 - " << endl;
classify();
cout << endl;
sample = -0.0702862, 0.065316, ..., -0.0279446, 0.0453012;
cout << "Unknown - " << endl;
classify();
cout << endl;
sample = -0.0789684, 0.0632067, ..., 0.0330486, 0.0117508;
cout << "Unknown - " << endl;
classify();
cout << endl;
sample = -0.0941284, 0.0542927, ..., 0.00855513, 0.00840678;
cout << "Unknown - " << endl;
classify();
// <- Test
return 0;
}