Add detector train implementation

apps/icf/icfdetector.cpp

@@ -1,22 +1,119 @@
 #include "icfdetector.hpp"
+#include "waldboost.hpp"
 
 #include <iostream>
 
+#include <sstream>
+using std::ostringstream;
+
 using std::vector;
 using std::string;
 
+#include <algorithm>
+using std::min;
+using std::max;
+
 #include <opencv2/core.hpp>
-using cv::Rect;
+#include <opencv2/imgproc.hpp>
+#include <opencv2/highgui.hpp>
+
+namespace cv
+{
+namespace adas
+{
+
+static bool overlap(const Rect& r, const vector<Rect>& gt)
+{
+    for( size_t i = 0; i < gt.size(); ++i )
+        if( (r & gt[i]).area() )
+            return true;
+    return false;
+}
 
 void ICFDetector::train(const vector<string>& image_filenames,
                         const vector< vector<Rect> >& labelling,
                         ICFDetectorParams params)
 {
-    std::cout << "train" << std::endl;
+    Size model_size(params.model_n_cols, params.model_n_rows);
+
+    vector<Mat> samples; /* positive samples + negative samples */
+    Mat sample, resized_sample;
+    int pos_count = 0;
+    for( size_t i = 0; i < image_filenames.size(); ++i, ++pos_count )
+    {
+        Mat img = imread(image_filenames[i]);
+        for( size_t j = 0; j < labelling[i].size(); ++j )
+        {
+            Rect r = labelling[i][j];
+            if( r.x > img.cols || r.y > img.rows )
+                continue;
+
+            sample = img.colRange(max(r.x, 0), min(r.width, img.cols))
+                        .rowRange(max(r.y, 0), min(r.height, img.rows));
 
+            resize(sample, resized_sample, model_size);
+
+            samples.push_back(resized_sample);
+        }
+    }
+
+    int neg_count = 0;
+    RNG rng;
+    for( size_t i = 0; i < image_filenames.size(); ++i, ++neg_count )
+    {
+        Mat img = imread(image_filenames[i]);
+        for( size_t j = 0; j < pos_count / image_filenames.size() + 1; ++j )
+        {
+            Rect r;
+            r.x = rng.uniform(0, img.cols);
+            r.width = rng.uniform(r.x + 1, img.cols);
+            r.y = rng.uniform(0, img.rows);
+            r.height = rng.uniform(r.y + 1, img.rows);
+
+            if( !overlap(r, labelling[i]) )
+            {
+                sample = img.colRange(r.x, r.width).rowRange(r.y, r.height);
+                resize(sample, resized_sample);
+                samples.push_back(resized_sample);
+                ++neg_count;
+            }
+        }
+    }
+
+    Mat_<int> labels(1, pos_count + neg_count);
+    for( size_t i = 0; i < pos_count; ++i)
+        labels(0, i) = 1;
+    for( size_t i = pos_count; i < pos_count + neg_count; ++i )
+        labels(0, i) = -1;
+
+    vector<Point3i> features = generateFeatures(model_size);
+    ACFFeatureEvaluator feature_evaluator(features);
+
+    Mat_<int> data(features.size(), samples.size());
+    Mat_<int> feature_col;
+
+    vector<Mat> channels;
+    for( size_t i = 0; i < samples.size(); ++i )
+    {
+        computeChannels(samples[i], channels);
+        feature_evaluator.setChannels(channels);
+        feature_evaluator.evaluateAll(feature_col);
+        for( int j = 0; j < feature_col.rows; ++j )
+            data(i, j) = feature_col(0, j);
+    }
+
+    WaldBoostParams wparams;
+    wparams.weak_count = params.weak_count;
+    wparams.alpha = 0.001;
+
+    WaldBoost waldboost(wparams);
+    waldboost.train(data, labels);
 }
 
 bool ICFDetector::save(const string& filename)
 {
     return true;
 }
+
+} /* namespace adas */
+} /* namespace cv */

apps/icf/icfdetector.hpp

@@ -47,6 +47,11 @@ the use of this software, even if advised of the possibility of such damage.
 
 #include <opencv2/core.hpp>
 
+namespace cv
+{
+namespace adas
+{
+
 struct ICFDetectorParams
 {
     int feature_count;
@@ -75,4 +80,6 @@ public:
     bool save(const std::string& filename);
 };
 
+} /* namespace adas */
+} /* namespace cv */
 #endif /* __OPENCV_ADAS_ICFDETECTOR_HPP__ */