Add icf feature evaluator

modules/xobjdetect/include/opencv2/xobjdetect.hpp

@@ -58,9 +58,9 @@ namespace xobjdetect
     channels — output array for computed channels
 
 */
-void computeChannels(InputArray image, std::vector<Mat>& channels);
+CV_EXPORTS void computeChannels(InputArray image, std::vector<Mat>& channels);
 
-class CV_EXPORTS ACFFeatureEvaluator : public Algorithm
+class CV_EXPORTS FeatureEvaluator : public Algorithm
 {
 public:
     /* Set channels for feature evaluation */
@@ -69,8 +69,6 @@ public:
     /* Set window position */
     virtual void setPosition(Size position) = 0;
 
-    virtual void assertChannels() = 0;
-
     /* Evaluate feature with given index for current channels
         and window position */
     virtual int evaluate(size_t feature_ind) const = 0;
@@ -81,23 +79,28 @@ public:
     */
     virtual void evaluateAll(OutputArray feature_values) const = 0;
 
+    virtual void assertChannels() = 0;
 };
 
-/* Construct evaluator, set features to evaluate */
-CV_EXPORTS Ptr<ACFFeatureEvaluator>
-createACFFeatureEvaluator(const std::vector<Point3i>& features);
+/* Construct feature evaluator, set features to evaluate
+   type can "icf" or "acf" */
+CV_EXPORTS Ptr<FeatureEvaluator>
+createFeatureEvaluator(const std::vector<std::vector<int> >& features,
+                       const std::string& type);
 
 /* Generate acf features
 
     window_size — size of window in which features should be evaluated
 
+    type — type of features, can be "icf" or "acf"
     count — number of features to generate.
     Max number of features is min(count, # possible distinct features)
 
 Returns vector of distinct acf features
 */
-std::vector<Point3i>
-generateFeatures(Size window_size, int count = INT_MAX);
+std::vector<std::vector<int> >
+generateFeatures(Size window_size, const std::string& type,
+                 int count = INT_MAX, int channel_count = 10);
 
 
 struct CV_EXPORTS WaldBoostParams
@@ -135,7 +138,7 @@ public:
     is from class +1
     */
     virtual float predict(
-        const Ptr<ACFFeatureEvaluator>& /*feature_evaluator*/) const
+        const Ptr<FeatureEvaluator>& /*feature_evaluator*/) const
     {return 0.0f;}
 
     /* Write WaldBoost to FileStorage */
@@ -207,7 +210,7 @@ public:
 
 private:
     Ptr<WaldBoost> waldboost_;
-    std::vector<Point3i> features_;
+    std::vector<std::vector<int> > features_;
     int model_n_rows_;
     int model_n_cols_;
 };

modules/xobjdetect/src/acffeature.cpp

@@ -54,75 +54,126 @@ namespace cv
 namespace xobjdetect
 {
 
-class ACFFeatureEvaluatorImpl : public ACFFeatureEvaluator
+static bool isNull(const Mat_<int> &m)
+{
+    bool null_data = true;
+    for( int row = 0; row < m.rows; ++row )
+    {
+        for( int col = 0; col < m.cols; ++col )
+            if( m.at<int>(row, col) )
+                null_data = false;
+    }
+    return null_data;
+}
+
+class FeatureEvaluatorImpl : public FeatureEvaluator
 {
 public:
-    ACFFeatureEvaluatorImpl(const vector<Point3i>& features):
+    FeatureEvaluatorImpl(const vector<vector<int> >& features):
         features_(features), channels_(), position_()
     {
         CV_Assert(features.size() > 0);
     }
 
-    virtual void setChannels(InputArrayOfArrays channels);
-    virtual void assertChannels();
-    virtual void setPosition(Size position);
-    virtual int evaluate(size_t feature_ind) const;
-    virtual void evaluateAll(OutputArray feature_values) const;
+    virtual void assertChannels()
+    {
+        bool null_data = true;
+        for( size_t i = 0; i < channels_.size(); ++i )
+            null_data &= isNull(channels_[i]);
+        CV_Assert(!null_data);
+    }
+
+    virtual void evaluateAll(OutputArray feature_values) const
+    {
+        Mat_<int> feature_vals(1, (int)features_.size());
+        for( int i = 0; i < (int)features_.size(); ++i )
+        {
+            feature_vals(0, i) = evaluate(i);
+        }
+        feature_values.assign(feature_vals);
+    }
 
-private:
+protected:
     /* Features to evaluate */
-    std::vector<Point3i> features_;
+    vector<vector<int> > features_;
     /* Channels for feature evaluation */
     std::vector<Mat> channels_;
     /* Channels window position */
     Size position_;
 };
 
-
-static bool isNull(const Mat_<int> &m)
+class ICFFeatureEvaluatorImpl : public FeatureEvaluatorImpl
 {
-    bool null_data = true;
-    for( int row = 0; row < m.rows; ++row )
+public:
+    ICFFeatureEvaluatorImpl(const vector<vector<int> >& features):
+        FeatureEvaluatorImpl(features)
     {
-        for( int col = 0; col < m.cols; ++col )
-            if( m.at<int>(row, col) )
-                null_data = false;
     }
-    return null_data;
-}
 
-void ACFFeatureEvaluatorImpl::assertChannels()
-{
-    bool null_data = true;
-    for( size_t i = 0; i < channels_.size(); ++i )
-        null_data &= isNull(channels_[i]);
-    CV_Assert(!null_data);
-}
+    virtual void setChannels(InputArrayOfArrays channels);
+    virtual void setPosition(Size position);
+    virtual int evaluate(size_t feature_ind) const;
+};
 
-void ACFFeatureEvaluatorImpl::setChannels(cv::InputArrayOfArrays channels)
+void ICFFeatureEvaluatorImpl::setChannels(InputArrayOfArrays channels)
 {
     channels_.clear();
     vector<Mat> ch;
     channels.getMatVector(ch);
     CV_Assert(ch.size() == 10);
 
-    /*int min_val = 100500, max_val = -1;
     for( size_t i = 0; i < ch.size(); ++i )
     {
         const Mat &channel = ch[i];
-        for( int row = 0; row < channel.rows; ++row )
-            for( int col = 0; col < channel.cols; ++col )
-            {
-                int val = (int)channel.at<float>(row, col);
-                if( val < min_val )
-                    min_val = val;
-                else if( val > max_val )
-                    max_val = val;
-            }
+        Mat integral_channel;
+        integral(channel, integral_channel, CV_32F);
+        Mat_<int> chan(integral_channel.rows, integral_channel.cols);
+        for( int row = 0; row < integral_channel.rows; ++row )
+            for( int col = 0; col < integral_channel.cols; ++col )
+                chan(row, col) = (int)integral_channel.at<float>(row, col);
+        channels_.push_back(chan.clone());
     }
+}
+
+void ICFFeatureEvaluatorImpl::setPosition(Size position)
+{
+    position_ = position;
+}
+
+int ICFFeatureEvaluatorImpl::evaluate(size_t feature_ind) const
+{
+    CV_Assert(channels_.size() == 10);
+    CV_Assert(feature_ind < features_.size());
+
+    const vector<int>& feature = features_[feature_ind];
+    int x = feature[0] + position_.height;
+    int y = feature[1] + position_.width;
+    int x_to = feature[2] + position_.height;
+    int y_to = feature[3] + position_.width;
+    int n = feature[4];
+    const Mat_<int>& ch = channels_[n];
+    return ch(y_to + 1, x_to + 1) - ch(y, x_to + 1) - ch(y_to + 1, x) + ch(y, x);
+}
+
+class ACFFeatureEvaluatorImpl : public FeatureEvaluatorImpl
+{
+public:
+    ACFFeatureEvaluatorImpl(const vector<vector<int> >& features):
+        FeatureEvaluatorImpl(features)
+    {
+    }
+
+    virtual void setChannels(InputArrayOfArrays channels);
+    virtual void setPosition(Size position);
+    virtual int evaluate(size_t feature_ind) const;
+};
 
-    cout << "SET " << min_val << " " << max_val << endl;
-    */
+void ACFFeatureEvaluatorImpl::setChannels(InputArrayOfArrays channels)
+{
+    channels_.clear();
+    vector<Mat> ch;
+    channels.getMatVector(ch);
+    CV_Assert(ch.size() == 10);
 
     for( size_t i = 0; i < ch.size(); ++i )
     {
@@ -135,11 +186,7 @@ void ACFFeatureEvaluatorImpl::setChannels(cv::InputArrayOfArrays channels)
                 int sum = 0;
                 for( int cell_row = row; cell_row < row + 4; ++cell_row )
                     for( int cell_col = col; cell_col < col + 4; ++cell_col )
-                    {
-                        //cout << channel.rows << " " << channel.cols << endl;
-                        //cout << cell_row << " " << cell_col << endl;
                         sum += (int)channel.at<float>(cell_row, cell_col);
-                    }
 
                 acf_channel(row / 4, col / 4) = sum;
             }
@@ -159,49 +206,66 @@ int ACFFeatureEvaluatorImpl::evaluate(size_t feature_ind) const
     CV_Assert(channels_.size() == 10);
     CV_Assert(feature_ind < features_.size());
 
-    Point3i feature = features_.at(feature_ind);
-    int x = feature.x;
-    int y = feature.y;
-    int n = feature.z;
+    const vector<int>& feature = features_[feature_ind];
+    int x = feature[0];
+    int y = feature[1];
+    int n = feature[2];
     return channels_[n].at<int>(y + position_.width, x + position_.height);
 }
 
-void ACFFeatureEvaluatorImpl::evaluateAll(OutputArray feature_values) const
+Ptr<FeatureEvaluator> createFeatureEvaluator(
+    const vector<vector<int> >& features, const std::string& type)
 {
-    Mat_<int> feature_vals(1, (int)features_.size());
-    for( int i = 0; i < (int)features_.size(); ++i )
-    {
-        feature_vals(0, i) = evaluate(i);
-    }
-    feature_values.assign(feature_vals);
+    FeatureEvaluator *evaluator = NULL;
+    if( type == "acf" )
+        evaluator = new ACFFeatureEvaluatorImpl(features);
+    else if( type == "icf" )
+        evaluator = new ICFFeatureEvaluatorImpl(features);
+    else
+        CV_Assert(false);
+
+    return Ptr<FeatureEvaluator>(evaluator);
 }
 
-Ptr<ACFFeatureEvaluator>
-createACFFeatureEvaluator(const vector<Point3i>& features)
-{
-    return Ptr<ACFFeatureEvaluator>(new ACFFeatureEvaluatorImpl(features));
-}
-
-vector<Point3i> generateFeatures(Size window_size, int count)
+vector<vector<int> > generateFeatures(Size window_size, const std::string& type,
+                                      int count, int channel_count)
 {
     CV_Assert(count > 0);
-    int cur_count = 0;
-    int max_count = window_size.width * window_size.height / 16;
-    count = min(count, max_count);
-    vector<Point3i> features;
-    for( int x = 0; x < window_size.width / 4; ++x )
+    vector<vector<int> > features;
+    if( type == "acf" )
     {
-        for( int y = 0; y < window_size.height / 4; ++y )
+        int cur_count = 0;
+        int max_count = window_size.width * window_size.height / 16;
+        count = min(count, max_count);
+        for( int x = 0; x < window_size.width / 4; ++x )
+            for( int y = 0; y < window_size.height / 4; ++y )
+                for( int n = 0; n < channel_count; ++n )
+                {
+                    int f[] = {x, y, n};
+                    vector<int> feature(f, f + sizeof(f) / sizeof(*f));
+                    features.push_back(feature);
+                    if( (cur_count += 1) == count )
+                        break;
+                }
+    }
+    else if( type == "icf" )
+    {
+        RNG rng;
+        for( int i = 0; i < count; ++i )
         {
-            /* Assume there are 10 channel types */
-            for( int n = 0; n < 10; ++n )
-            {
-                features.push_back(Point3i(x, y, n));
-                if( (cur_count += 1) == count )
-                    break;
-            }
+            int x = rng.uniform(0, window_size.width - 1);
+            int y = rng.uniform(0, window_size.height - 1);
+            int x_to = rng.uniform(x, window_size.width - 1);
+            int y_to = rng.uniform(y, window_size.height - 1);
+            int n = rng.uniform(0, channel_count - 1);
+            int f[] = {x, y, x_to, y_to, n};
+            vector<int> feature(f, f + sizeof(f) / sizeof(*f));
+            features.push_back(feature);
         }
     }
+    else
+        CV_Assert(false);
+
     return features;
 }
 

modules/xobjdetect/src/icfdetector.cpp

@@ -118,8 +118,9 @@ void ICFDetector::train(const String& pos_path,
     for( int i = pos_count; i < pos_count + neg_count; ++i )
         labels(0, i) = -1;
 
-    vector<Point3i> features = generateFeatures(model_size);
-    Ptr<ACFFeatureEvaluator> feature_evaluator = createACFFeatureEvaluator(features);
+    vector<vector<int> > features = generateFeatures(model_size, "icf",
+        params.feature_count);
+    Ptr<FeatureEvaluator> evaluator = createFeatureEvaluator(features, "icf");
 
     Mat_<int> data = Mat_<int>::zeros((int)features.size(), (int)samples.size());
     Mat_<int> feature_col(1, (int)samples.size());
@@ -129,9 +130,9 @@ void ICFDetector::train(const String& pos_path,
     {
         cout << setw(6) << i << "/" << samples.size() << "\r";
         computeChannels(samples[i], channels);
-        feature_evaluator->setChannels(channels);
-        //feature_evaluator->assertChannels();
-        feature_evaluator->evaluateAll(feature_col);
+        evaluator->setChannels(channels);
+        //evaluator->assertChannels();
+        evaluator->evaluateAll(feature_col);
 
         CV_Assert(feature_col.cols == (int)features.size());
 
@@ -180,7 +181,7 @@ void ICFDetector::read(const FileNode& node)
     node["waldboost"] >> *waldboost_;
     FileNode features = node["features"];
     features_.clear();
-    Point3i p;
+    vector<int> p;
     for( FileNodeIterator n = features.begin(); n != features.end(); ++n )
     {
         (*n) >> p;
@@ -196,7 +197,7 @@ void ICFDetector::detect(const Mat& img, vector<Rect>& objects,
     float scale_to = max(model_n_cols_ / (float)minSize.width,
                          model_n_rows_ / (float)minSize.height);
     objects.clear();
-    Ptr<ACFFeatureEvaluator> evaluator = createACFFeatureEvaluator(features_);
+    Ptr<FeatureEvaluator> evaluator = createFeatureEvaluator(features_, "icf");
     Mat rescaled_image;
     int step = 8;
     vector<Mat> channels;

modules/xobjdetect/src/waldboost.cpp

@@ -67,7 +67,7 @@ public:
                                    const Mat& labels);
 
     virtual float predict(
-        const Ptr<ACFFeatureEvaluator>& feature_evaluator) const;
+        const Ptr<FeatureEvaluator>& feature_evaluator) const;
 
     virtual void write(FileStorage& fs) const;
 
@@ -299,7 +299,7 @@ vector<int> WaldBoostImpl::train(const Mat& data_, const Mat& labels_)
 }
 
 float WaldBoostImpl::predict(
-    const Ptr<ACFFeatureEvaluator>& feature_evaluator) const
+    const Ptr<FeatureEvaluator>& feature_evaluator) const
 {
     float trace = 0;
     CV_Assert(stumps_.size() == thresholds_.size());