refactor CUDA HOG algorithm:

use abstract interface with hidden implementation

modules/cudaobjdetect/include/opencv2/cudaobjdetect.hpp

@@ -65,32 +65,24 @@ namespace cv { namespace cuda {
 // HOG (Histogram-of-Oriented-Gradients) Descriptor and Object Detector
 //
 
-struct CV_EXPORTS HOGConfidence
-{
-   double scale;
-   std::vector<Point> locations;
-   std::vector<double> confidences;
-   std::vector<double> part_scores[4];
-};
-
 /** @brief The class implements Histogram of Oriented Gradients (@cite Dalal2005) object detector.
 
-Interfaces of all methods are kept similar to the CPU HOG descriptor and detector analogues as much
-as possible.
-
 @note
-   -   An example applying the HOG descriptor for people detection can be found at
+    -   An example applying the HOG descriptor for people detection can be found at
         opencv_source_code/samples/cpp/peopledetect.cpp
     -   A CUDA example applying the HOG descriptor for people detection can be found at
         opencv_source_code/samples/gpu/hog.cpp
     -   (Python) An example applying the HOG descriptor for people detection can be found at
         opencv_source_code/samples/python2/peopledetect.py
  */
-struct CV_EXPORTS HOGDescriptor
+class CV_EXPORTS HOG : public cv::Algorithm
 {
-    enum { DEFAULT_WIN_SIGMA = -1 };
-    enum { DEFAULT_NLEVELS = 64 };
-    enum { DESCR_FORMAT_ROW_BY_ROW, DESCR_FORMAT_COL_BY_COL };
+public:
+    enum
+    {
+        DESCR_FORMAT_ROW_BY_ROW,
+        DESCR_FORMAT_COL_BY_COL
+    };
 
     /** @brief Creates the HOG descriptor and detector.
 
@@ -99,132 +91,105 @@ struct CV_EXPORTS HOGDescriptor
     @param block_stride Block stride. It must be a multiple of cell size.
     @param cell_size Cell size. Only (8, 8) is supported for now.
     @param nbins Number of bins. Only 9 bins per cell are supported for now.
-    @param win_sigma Gaussian smoothing window parameter.
-    @param threshold_L2hys L2-Hys normalization method shrinkage.
-    @param gamma_correction Flag to specify whether the gamma correction preprocessing is required or
-    not.
-    @param nlevels Maximum number of detection window increases.
      */
-    HOGDescriptor(Size win_size=Size(64, 128), Size block_size=Size(16, 16),
-                  Size block_stride=Size(8, 8), Size cell_size=Size(8, 8),
-                  int nbins=9, double win_sigma=DEFAULT_WIN_SIGMA,
-                  double threshold_L2hys=0.2, bool gamma_correction=true,
-                  int nlevels=DEFAULT_NLEVELS);
+    static Ptr<HOG> create(Size win_size = Size(64, 128),
+                           Size block_size = Size(16, 16),
+                           Size block_stride = Size(8, 8),
+                           Size cell_size = Size(8, 8),
+                           int nbins = 9);
+
+    //! Gaussian smoothing window parameter.
+    virtual void setWinSigma(double win_sigma) = 0;
+    virtual double getWinSigma() const = 0;
+
+    //! L2-Hys normalization method shrinkage.
+    virtual void setL2HysThreshold(double threshold_L2hys) = 0;
+    virtual double getL2HysThreshold() const = 0;
+
+    //! Flag to specify whether the gamma correction preprocessing is required or not.
+    virtual void setGammaCorrection(bool gamma_correction) = 0;
+    virtual bool getGammaCorrection() const = 0;
+
+    //! Maximum number of detection window increases.
+    virtual void setNumLevels(int nlevels) = 0;
+    virtual int getNumLevels() const = 0;
+
+    //! Threshold for the distance between features and SVM classifying plane.
+    //! Usually it is 0 and should be specfied in the detector coefficients (as the last free
+    //! coefficient). But if the free coefficient is omitted (which is allowed), you can specify it
+    //! manually here.
+    virtual void setHitThreshold(double hit_threshold) = 0;
+    virtual double getHitThreshold() const = 0;
+
+    //! Window stride. It must be a multiple of block stride.
+    virtual void setWinStride(Size win_stride) = 0;
+    virtual Size getWinStride() const = 0;
+
+    //! Coefficient of the detection window increase.
+    virtual void setScaleFactor(double scale0) = 0;
+    virtual double getScaleFactor() const = 0;
+
+    //! Coefficient to regulate the similarity threshold. When detected, some
+    //! objects can be covered by many rectangles. 0 means not to perform grouping.
+    //! See groupRectangles.
+    virtual void setGroupThreshold(int group_threshold) = 0;
+    virtual int getGroupThreshold() const = 0;
+
+    //! Descriptor storage format:
+    //!   - **DESCR_FORMAT_ROW_BY_ROW** - Row-major order.
+    //!   - **DESCR_FORMAT_COL_BY_COL** - Column-major order.
+    virtual void setDescriptorFormat(int descr_format) = 0;
+    virtual int getDescriptorFormat() const = 0;
 
     /** @brief Returns the number of coefficients required for the classification.
      */
-    size_t getDescriptorSize() const;
+    virtual size_t getDescriptorSize() const = 0;
+
     /** @brief Returns the block histogram size.
-    */
-    size_t getBlockHistogramSize() const;
+     */
+    virtual size_t getBlockHistogramSize() const = 0;
 
     /** @brief Sets coefficients for the linear SVM classifier.
-    */
-    void setSVMDetector(const std::vector<float>& detector);
+     */
+    virtual void setSVMDetector(InputArray detector) = 0;
 
-    /** @brief Returns coefficients of the classifier trained for people detection (for default window size).
-    */
-    static std::vector<float> getDefaultPeopleDetector();
-    /** @brief Returns coefficients of the classifier trained for people detection (for 48x96 windows).
-    */
-    static std::vector<float> getPeopleDetector48x96();
-    /** @brief Returns coefficients of the classifier trained for people detection (for 64x128 windows).
-    */
-    static std::vector<float> getPeopleDetector64x128();
+    /** @brief Returns coefficients of the classifier trained for people detection.
+     */
+    virtual Mat getDefaultPeopleDetector() const = 0;
 
     /** @brief Performs object detection without a multi-scale window.
 
     @param img Source image. CV_8UC1 and CV_8UC4 types are supported for now.
     @param found_locations Left-top corner points of detected objects boundaries.
-    @param hit_threshold Threshold for the distance between features and SVM classifying plane.
-    Usually it is 0 and should be specfied in the detector coefficients (as the last free
-    coefficient). But if the free coefficient is omitted (which is allowed), you can specify it
-    manually here.
-    @param win_stride Window stride. It must be a multiple of block stride.
-    @param padding Mock parameter to keep the CPU interface compatibility. It must be (0,0).
+    @param confidences Optional output array for confidences.
      */
-    void detect(const GpuMat& img, std::vector<Point>& found_locations,
-                double hit_threshold=0, Size win_stride=Size(),
-                Size padding=Size());
+    virtual void detect(InputArray img,
+                        std::vector<Point>& found_locations,
+                        std::vector<double>* confidences = NULL) = 0;
 
     /** @brief Performs object detection with a multi-scale window.
 
     @param img Source image. See cuda::HOGDescriptor::detect for type limitations.
     @param found_locations Detected objects boundaries.
+    @param confidences Optional output array for confidences.
     @param hit_threshold Threshold for the distance between features and SVM classifying plane. See
     cuda::HOGDescriptor::detect for details.
     @param win_stride Window stride. It must be a multiple of block stride.
     @param padding Mock parameter to keep the CPU interface compatibility. It must be (0,0).
-    @param scale0 Coefficient of the detection window increase.
-    @param group_threshold Coefficient to regulate the similarity threshold. When detected, some
-    objects can be covered by many rectangles. 0 means not to perform grouping. See groupRectangles .
      */
-    void detectMultiScale(const GpuMat& img, std::vector<Rect>& found_locations,
-                          double hit_threshold=0, Size win_stride=Size(),
-                          Size padding=Size(), double scale0=1.05,
-                          int group_threshold=2);
-
-    void computeConfidence(const GpuMat& img, std::vector<Point>& hits, double hit_threshold,
-                                                Size win_stride, Size padding, std::vector<Point>& locations, std::vector<double>& confidences);
-
-    void computeConfidenceMultiScale(const GpuMat& img, std::vector<Rect>& found_locations,
-                                                                    double hit_threshold, Size win_stride, Size padding,
-                                                                    std::vector<HOGConfidence> &conf_out, int group_threshold);
+    virtual void detectMultiScale(InputArray img,
+                                  std::vector<Rect>& found_locations,
+                                  std::vector<double>* confidences = NULL) = 0;
 
     /** @brief Returns block descriptors computed for the whole image.
 
     @param img Source image. See cuda::HOGDescriptor::detect for type limitations.
-    @param win_stride Window stride. It must be a multiple of block stride.
     @param descriptors 2D array of descriptors.
-    @param descr_format Descriptor storage format:
-    -   **DESCR_FORMAT_ROW_BY_ROW** - Row-major order.
-    -   **DESCR_FORMAT_COL_BY_COL** - Column-major order.
-
-    The function is mainly used to learn the classifier.
+    @param stream CUDA stream.
      */
-    void getDescriptors(const GpuMat& img, Size win_stride,
-                        GpuMat& descriptors,
-                        int descr_format=DESCR_FORMAT_COL_BY_COL);
-
-    Size win_size;
-    Size block_size;
-    Size block_stride;
-    Size cell_size;
-    int nbins;
-    double win_sigma;
-    double threshold_L2hys;
-    bool gamma_correction;
-    int nlevels;
-
-protected:
-    void computeBlockHistograms(const GpuMat& img);
-    void computeGradient(const GpuMat& img, GpuMat& grad, GpuMat& qangle);
-
-    double getWinSigma() const;
-    bool checkDetectorSize() const;
-
-    static int numPartsWithin(int size, int part_size, int stride);
-    static Size numPartsWithin(Size size, Size part_size, Size stride);
-
-    // Coefficients of the separating plane
-    float free_coef;
-    GpuMat detector;
-
-    // Results of the last classification step
-    GpuMat labels, labels_buf;
-    Mat labels_host;
-
-    // Results of the last histogram evaluation step
-    GpuMat block_hists, block_hists_buf;
-
-    // Gradients conputation results
-    GpuMat grad, qangle, grad_buf, qangle_buf;
-
-    // returns subbuffer with required size, reallocates buffer if nessesary.
-    static GpuMat getBuffer(const Size& sz, int type, GpuMat& buf);
-    static GpuMat getBuffer(int rows, int cols, int type, GpuMat& buf);
-
-    std::vector<GpuMat> image_scales;
+    virtual void compute(InputArray img,
+                         OutputArray descriptors,
+                         Stream& stream = Stream::Null()) = 0;
 };
 
 //

modules/cudaobjdetect/perf/perf_objdetect.cpp

@@ -71,10 +71,10 @@ PERF_TEST_P(Image, ObjDetect_HOG,
         const cv::cuda::GpuMat d_img(img);
         std::vector<cv::Rect> gpu_found_locations;
 
-        cv::cuda::HOGDescriptor d_hog;
-        d_hog.setSVMDetector(cv::cuda::HOGDescriptor::getDefaultPeopleDetector());
+        cv::Ptr<cv::cuda::HOG> d_hog = cv::cuda::HOG::create();
+        d_hog->setSVMDetector(d_hog->getDefaultPeopleDetector());
 
-        TEST_CYCLE() d_hog.detectMultiScale(d_img, gpu_found_locations);
+        TEST_CYCLE() d_hog->detectMultiScale(d_img, gpu_found_locations);
 
         SANITY_CHECK(gpu_found_locations);
     }
@@ -82,8 +82,10 @@ PERF_TEST_P(Image, ObjDetect_HOG,
     {
         std::vector<cv::Rect> cpu_found_locations;
 
+        cv::Ptr<cv::cuda::HOG> d_hog = cv::cuda::HOG::create();
+
         cv::HOGDescriptor hog;
-        hog.setSVMDetector(cv::cuda::HOGDescriptor::getDefaultPeopleDetector());
+        hog.setSVMDetector(d_hog->getDefaultPeopleDetector());
 
         TEST_CYCLE() hog.detectMultiScale(img, cpu_found_locations);
 

modules/cudaobjdetect/test/test_objdetect.cpp

@@ -48,9 +48,10 @@ using namespace cvtest;
 
 //#define DUMP
 
-struct HOG : testing::TestWithParam<cv::cuda::DeviceInfo>, cv::cuda::HOGDescriptor
+struct HOG : testing::TestWithParam<cv::cuda::DeviceInfo>
 {
     cv::cuda::DeviceInfo devInfo;
+    cv::Ptr<cv::cuda::HOG> hog;
 
 #ifdef DUMP
     std::ofstream f;
@@ -69,23 +70,13 @@ struct HOG : testing::TestWithParam<cv::cuda::DeviceInfo>, cv::cuda::HOGDescript
         devInfo = GetParam();
 
         cv::cuda::setDevice(devInfo.deviceID());
+
+        hog = cv::cuda::HOG::create();
     }
 
 #ifdef DUMP
-    void dump(const cv::Mat& blockHists, const std::vector<cv::Point>& locations)
+    void dump(const std::vector<cv::Point>& locations)
     {
-        f.write((char*)&blockHists.rows, sizeof(blockHists.rows));
-        f.write((char*)&blockHists.cols, sizeof(blockHists.cols));
-
-        for (int i = 0; i < blockHists.rows; ++i)
-        {
-            for (int j = 0; j < blockHists.cols; ++j)
-            {
-                float val = blockHists.at<float>(i, j);
-                f.write((char*)&val, sizeof(val));
-            }
-        }
-
         int nlocations = locations.size();
         f.write((char*)&nlocations, sizeof(nlocations));
 
@@ -93,21 +84,18 @@ struct HOG : testing::TestWithParam<cv::cuda::DeviceInfo>, cv::cuda::HOGDescript
             f.write((char*)&locations[i], sizeof(locations[i]));
     }
 #else
-    void compare(const cv::Mat& blockHists, const std::vector<cv::Point>& locations)
+    void compare(const std::vector<cv::Point>& locations)
     {
+        // skip block_hists check
         int rows, cols;
         f.read((char*)&rows, sizeof(rows));
         f.read((char*)&cols, sizeof(cols));
-        ASSERT_EQ(rows, blockHists.rows);
-        ASSERT_EQ(cols, blockHists.cols);
-
-        for (int i = 0; i < blockHists.rows; ++i)
+        for (int i = 0; i < rows; ++i)
         {
-            for (int j = 0; j < blockHists.cols; ++j)
+            for (int j = 0; j < cols; ++j)
             {
                 float val;
                 f.read((char*)&val, sizeof(val));
-                ASSERT_NEAR(val, blockHists.at<float>(i, j), 1e-3);
             }
         }
 
@@ -126,54 +114,41 @@ struct HOG : testing::TestWithParam<cv::cuda::DeviceInfo>, cv::cuda::HOGDescript
 
     void testDetect(const cv::Mat& img)
     {
-        gamma_correction = false;
-        setSVMDetector(cv::cuda::HOGDescriptor::getDefaultPeopleDetector());
+        hog->setGammaCorrection(false);
+        hog->setSVMDetector(hog->getDefaultPeopleDetector());
 
         std::vector<cv::Point> locations;
 
         // Test detect
-        detect(loadMat(img), locations, 0);
+        hog->detect(loadMat(img), locations);
 
 #ifdef DUMP
-        dump(cv::Mat(block_hists), locations);
+        dump(locations);
 #else
-        compare(cv::Mat(block_hists), locations);
+        compare(locations);
 #endif
 
         // Test detect on smaller image
         cv::Mat img2;
         cv::resize(img, img2, cv::Size(img.cols / 2, img.rows / 2));
-        detect(loadMat(img2), locations, 0);
+        hog->detect(loadMat(img2), locations);
 
 #ifdef DUMP
-        dump(cv::Mat(block_hists), locations);
+        dump(locations);
 #else
-        compare(cv::Mat(block_hists), locations);
+        compare(locations);
 #endif
 
         // Test detect on greater image
         cv::resize(img, img2, cv::Size(img.cols * 2, img.rows * 2));
-        detect(loadMat(img2), locations, 0);
+        hog->detect(loadMat(img2), locations);
 
 #ifdef DUMP
-        dump(cv::Mat(block_hists), locations);
+        dump(locations);
 #else
-        compare(cv::Mat(block_hists), locations);
+        compare(locations);
 #endif
     }
-
-    // Does not compare border value, as interpolation leads to delta
-    void compare_inner_parts(cv::Mat d1, cv::Mat d2)
-    {
-        for (int i = 1; i < blocks_per_win_y - 1; ++i)
-            for (int j = 1; j < blocks_per_win_x - 1; ++j)
-                for (int k = 0; k < block_hist_size; ++k)
-                {
-                    float a = d1.at<float>(0, (i * blocks_per_win_x + j) * block_hist_size);
-                    float b = d2.at<float>(0, (i * blocks_per_win_x + j) * block_hist_size);
-                    ASSERT_FLOAT_EQ(a, b);
-                }
-    }
 };
 
 // desabled while resize does not fixed
@@ -182,13 +157,8 @@ CUDA_TEST_P(HOG, DISABLED_Detect)
     cv::Mat img_rgb = readImage("hog/road.png");
     ASSERT_FALSE(img_rgb.empty());
 
-#ifdef DUMP
     f.open((std::string(cvtest::TS::ptr()->get_data_path()) + "hog/expected_output.bin").c_str(), std::ios_base::binary);
     ASSERT_TRUE(f.is_open());
-#else
-    f.open((std::string(cvtest::TS::ptr()->get_data_path()) + "hog/expected_output.bin").c_str(), std::ios_base::binary);
-    ASSERT_TRUE(f.is_open());
-#endif
 
     // Test on color image
     cv::Mat img;
@@ -198,8 +168,6 @@ CUDA_TEST_P(HOG, DISABLED_Detect)
     // Test on gray image
     cv::cvtColor(img_rgb, img, cv::COLOR_BGR2GRAY);
     testDetect(img);
-
-    f.close();
 }
 
 CUDA_TEST_P(HOG, GetDescriptors)
@@ -216,8 +184,14 @@ CUDA_TEST_P(HOG, GetDescriptors)
 
     // Convert train images into feature vectors (train table)
     cv::cuda::GpuMat descriptors, descriptors_by_cols;
-    getDescriptors(d_img, win_size, descriptors, DESCR_FORMAT_ROW_BY_ROW);
-    getDescriptors(d_img, win_size, descriptors_by_cols, DESCR_FORMAT_COL_BY_COL);
+
+    hog->setWinStride(Size(64, 128));
+
+    hog->setDescriptorFormat(cv::cuda::HOG::DESCR_FORMAT_ROW_BY_ROW);
+    hog->compute(d_img, descriptors);
+
+    hog->setDescriptorFormat(cv::cuda::HOG::DESCR_FORMAT_COL_BY_COL);
+    hog->compute(d_img, descriptors_by_cols);
 
     // Check size of the result train table
     wins_per_img_x = 3;
@@ -242,48 +216,6 @@ CUDA_TEST_P(HOG, GetDescriptors)
                     ASSERT_EQ(l[(y * blocks_per_win_x + x) * block_hist_size + k],
                               r[(x * blocks_per_win_y + y) * block_hist_size + k]);
     }
-
-    /* Now we want to extract the same feature vectors, but from single images. NOTE: results will
-    be defferent, due to border values interpolation. Using of many small images is slower, however we
-    wont't call getDescriptors and will use computeBlockHistograms instead of. computeBlockHistograms
-    works good, it can be checked in the gpu_hog sample */
-
-    img_rgb = readImage("hog/positive1.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    // Everything is fine with interpolation for left top subimage
-    ASSERT_EQ(0.0, cv::norm((cv::Mat)block_hists, (cv::Mat)descriptors.rowRange(0, 1)));
-
-    img_rgb = readImage("hog/positive2.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    compare_inner_parts(cv::Mat(block_hists), cv::Mat(descriptors.rowRange(1, 2)));
-
-    img_rgb = readImage("hog/negative1.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    compare_inner_parts(cv::Mat(block_hists), cv::Mat(descriptors.rowRange(2, 3)));
-
-    img_rgb = readImage("hog/negative2.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    compare_inner_parts(cv::Mat(block_hists), cv::Mat(descriptors.rowRange(3, 4)));
-
-    img_rgb = readImage("hog/positive3.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    compare_inner_parts(cv::Mat(block_hists), cv::Mat(descriptors.rowRange(4, 5)));
-
-    img_rgb = readImage("hog/negative3.png");
-    ASSERT_TRUE(!img_rgb.empty());
-    cv::cvtColor(img_rgb, img, cv::COLOR_BGR2BGRA);
-    computeBlockHistograms(cv::cuda::GpuMat(img));
-    compare_inner_parts(cv::Mat(block_hists), cv::Mat(descriptors.rowRange(5, 6)));
 }
 
 INSTANTIATE_TEST_CASE_P(CUDA_ObjDetect, HOG, ALL_DEVICES);
@@ -310,12 +242,12 @@ CUDA_TEST_P(CalTech, HOG)
     cv::cuda::GpuMat d_img(img);
     cv::Mat markedImage(img.clone());
 
-    cv::cuda::HOGDescriptor d_hog;
-    d_hog.setSVMDetector(cv::cuda::HOGDescriptor::getDefaultPeopleDetector());
-    d_hog.nlevels = d_hog.nlevels + 32;
+    cv::Ptr<cv::cuda::HOG> d_hog = cv::cuda::HOG::create();
+    d_hog->setSVMDetector(d_hog->getDefaultPeopleDetector());
+    d_hog->setNumLevels(d_hog->getNumLevels() + 32);
 
     std::vector<cv::Rect> found_locations;
-    d_hog.detectMultiScale(d_img, found_locations);
+    d_hog->detectMultiScale(d_img, found_locations);
 
 #if defined (LOG_CASCADE_STATISTIC)
     for (int i = 0; i < (int)found_locations.size(); i++)
@@ -326,7 +258,8 @@ CUDA_TEST_P(CalTech, HOG)
         cv::rectangle(markedImage, r , CV_RGB(255, 0, 0));
     }
 
-    cv::imshow("Res", markedImage); cv::waitKey();
+    cv::imshow("Res", markedImage);
+    cv::waitKey();
 #endif
 }
 

samples/gpu/hog.cpp

@@ -244,19 +244,13 @@ void App::run()
     Size win_size(args.win_width, args.win_width * 2); //(64, 128) or (48, 96)
     Size win_stride(args.win_stride_width, args.win_stride_height);
 
+    cv::Ptr<cv::cuda::HOG> gpu_hog = cv::cuda::HOG::create(win_size);
+    cv::HOGDescriptor cpu_hog(win_size, Size(16, 16), Size(8, 8), Size(8, 8), 9);
+
     // Create HOG descriptors and detectors here
-    vector<float> detector;
-    if (win_size == Size(64, 128))
-        detector = cv::cuda::HOGDescriptor::getPeopleDetector64x128();
-    else
-        detector = cv::cuda::HOGDescriptor::getPeopleDetector48x96();
-
-    cv::cuda::HOGDescriptor gpu_hog(win_size, Size(16, 16), Size(8, 8), Size(8, 8), 9,
-                                   cv::cuda::HOGDescriptor::DEFAULT_WIN_SIGMA, 0.2, gamma_corr,
-                                   cv::cuda::HOGDescriptor::DEFAULT_NLEVELS);
-    cv::HOGDescriptor cpu_hog(win_size, Size(16, 16), Size(8, 8), Size(8, 8), 9, 1, -1,
-                              HOGDescriptor::L2Hys, 0.2, gamma_corr, cv::HOGDescriptor::DEFAULT_NLEVELS);
-    gpu_hog.setSVMDetector(detector);
+    Mat detector = gpu_hog->getDefaultPeopleDetector();
+
+    gpu_hog->setSVMDetector(detector);
     cpu_hog.setSVMDetector(detector);
 
     while (running)
@@ -307,9 +301,6 @@ void App::run()
             else img = img_aux;
             img_to_show = img;
 
-            gpu_hog.nlevels = nlevels;
-            cpu_hog.nlevels = nlevels;
-
             vector<Rect> found;
 
             // Perform HOG classification
@@ -317,11 +308,19 @@ void App::run()
             if (use_gpu)
             {
                 gpu_img.upload(img);
-                gpu_hog.detectMultiScale(gpu_img, found, hit_threshold, win_stride,
-                                         Size(0, 0), scale, gr_threshold);
+                gpu_hog->setNumLevels(nlevels);
+                gpu_hog->setHitThreshold(hit_threshold);
+                gpu_hog->setWinStride(win_stride);
+                gpu_hog->setScaleFactor(scale);
+                gpu_hog->setGroupThreshold(gr_threshold);
+                gpu_hog->detectMultiScale(gpu_img, found);
             }
-            else cpu_hog.detectMultiScale(img, found, hit_threshold, win_stride,
+            else
+            {
+                cpu_hog.nlevels = nlevels;
+                cpu_hog.detectMultiScale(img, found, hit_threshold, win_stride,
                                           Size(0, 0), scale, gr_threshold);
+            }
             hogWorkEnd();
 
             // Draw positive classified windows