Merge pull request #1390 from sovrasov:saliency_update

Author: vpisarev

modules/saliency/include/opencv2/saliency/saliencyBaseClasses.hpp

@@ -67,11 +67,6 @@ class CV_EXPORTS_W Saliency : public virtual Algorithm
    */
   virtual ~Saliency();
 
-  /**
-   * \brief Create Saliency by saliency type.
-   */
-  static Ptr<Saliency> create( const String& saliencyType );
-
   /**
    * \brief Compute the saliency
    * \param image        The image.
@@ -80,12 +75,6 @@ class CV_EXPORTS_W Saliency : public virtual Algorithm
    */
   CV_WRAP bool computeSaliency( InputArray image, OutputArray saliencyMap );
 
-  /**
-   * \brief Get the name of the specific saliency type
-   * \return The name of the tracker initializer
-   */
-  CV_WRAP String getClassName() const;
-
  protected:
 
   virtual bool computeSaliencyImpl( InputArray image, OutputArray saliencyMap ) = 0;

modules/saliency/include/opencv2/saliency/saliencySpecializedClasses.hpp

@@ -234,15 +234,28 @@ class CV_EXPORTS_W MotionSaliencyBinWangApr2014 : public MotionSaliency
   bool templateOrdering();
   bool templateReplacement( const Mat& finalBFMask, const Mat& image );
 
+  // Decision threshold adaptation and Activity control function
+  bool activityControl(const Mat& current_noisePixelsMask);
+  bool decisionThresholdAdaptation();
+
   // changing structure
   std::vector<Ptr<Mat> > backgroundModel;// The vector represents the background template T0---TK of reference paper.
   // Matrices are two-channel matrix. In the first layer there are the B (background value)
   // for each pixel. In the second layer, there are the C (efficacy) value for each pixel
   Mat potentialBackground;// Two channel Matrix. For each pixel, in the first level there are the Ba value (potential background value)
                           // and in the secon level there are the Ca value, the counter for each potential value.
-  Mat epslonPixelsValue;  // epslon threshold
+  Mat epslonPixelsValue;// epslon threshold
+
+  Mat activityPixelsValue;// Activity level of each pixel
+
+  //vector<Mat> noisePixelMask; // We define a ‘noise-pixel’ as a pixel that has been classified as a foreground pixel during the full resolution
+  Mat noisePixelMask;// We define a ‘noise-pixel’ as a pixel that has been classified as a foreground pixel during the full resolution
+  //detection process,however, after the low resolution detection, it has become a
+  // background pixel. The matrix is  two-channel matrix. In the first layer there is the mask ( the identified noise-pixels are set to 1 while other pixels are 0)
+  // for each pixel. In the second layer, there is the value of activity level A for each pixel.
 
   //fixed parameter
+  bool activityControlFlag;
   bool neighborhoodCheck;
   int N_DS;// Number of template to be downsampled and used in lowResolutionDetection function
   CV_PROP_RW int imageWidth;// Width of input image
@@ -257,6 +270,13 @@ class CV_EXPORTS_W MotionSaliencyBinWangApr2014 : public MotionSaliency
             // long-term template, regardless of any subsequent background changes. A relatively large (eg gamma=3) will
             //restrain the generation of ghosts.
 
+  uchar Ainc;// Activity Incrementation;
+  int Bmax;// Upper-bound value for pixel activity
+  int Bth;// Max activity threshold
+  int Binc, Bdec;// Threshold for pixel-level decision threshold (epslon) adaptation
+  float deltaINC, deltaDEC;// Increment-decrement value for epslon adaptation
+  int epslonMIN, epslonMAX;// Range values for epslon threshold
+
 };
 
 /************************************ Specific Objectness Specialized Classes ************************************/
@@ -417,38 +437,38 @@ class CV_EXPORTS_W ObjectnessBING : public Objectness
   int _Clr;//
   static const char* _clrName[3];
 
-  // Names and paths to read model and to store results
+// Names and paths to read model and to store results
   std::string _modelName, _bbResDir, _trainingPath, _resultsDir;
 
   std::vector<int> _svmSzIdxs;// Indexes of active size. It's equal to _svmFilters.size() and _svmReW1f.rows
   Mat _svmFilter;// Filters learned at stage I, each is a _H by _W CV_32F matrix
   FilterTIG _tigF;// TIG filter
   Mat _svmReW1f;// Re-weight parameters learned at stage II.
 
-  // List of the rectangles' objectness value, in the same order as
-  // the  vector<Vec4i> objectnessBoundingBox returned by the algorithm (in computeSaliencyImpl function)
+// List of the rectangles' objectness value, in the same order as
+// the  vector<Vec4i> objectnessBoundingBox returned by the algorithm (in computeSaliencyImpl function)
   std::vector<float> objectnessValues;
 
 private:
-  // functions
+// functions
 
   inline static float LoG( float x, float y, float delta )
   {
     float d = - ( x * x + y * y ) / ( 2 * delta * delta );
     return -1.0f / ( (float) ( CV_PI ) * pow( delta, 4 ) ) * ( 1 + d ) * exp( d );
   }  // Laplacian of Gaussian
 
-  // Read matrix from binary file
+// Read matrix from binary file
   static bool matRead( const std::string& filename, Mat& M );
 
   void setColorSpace( int clr = MAXBGR );
 
-  // Load trained model.
+// Load trained model.
   int loadTrainedModel( std::string modelName = "" );// Return -1, 0, or 1 if partial, none, or all loaded
 
-  // Get potential bounding boxes, each of which is represented by a Vec4i for (minX, minY, maxX, maxY).
-  // The trained model should be prepared before calling this function: loadTrainedModel() or trainStageI() + trainStageII().
-  // Use numDet to control the final number of proposed bounding boxes, and number of per size (scale and aspect ratio)
+// Get potential bounding boxes, each of which is represented by a Vec4i for (minX, minY, maxX, maxY).
+// The trained model should be prepared before calling this function: loadTrainedModel() or trainStageI() + trainStageII().
+// Use numDet to control the final number of proposed bounding boxes, and number of per size (scale and aspect ratio)
   void getObjBndBoxes( Mat &img3u, ValStructVec<float, Vec4i> &valBoxes, int numDetPerSize = 120 );
   void getObjBndBoxesForSingleImage( Mat img, ValStructVec<float, Vec4i> &boxes, int numDetPerSize );
 
@@ -460,7 +480,7 @@ class CV_EXPORTS_W ObjectnessBING : public Objectness
   void predictBBoxSI( Mat &mag3u, ValStructVec<float, Vec4i> &valBoxes, std::vector<int> &sz, int NUM_WIN_PSZ = 100, bool fast = true );
   void predictBBoxSII( ValStructVec<float, Vec4i> &valBoxes, const std::vector<int> &sz );
 
-  // Calculate the image gradient: center option as in VLFeat
+// Calculate the image gradient: center option as in VLFeat
   void gradientMag( Mat &imgBGR3u, Mat &mag1u );
 
   static void gradientRGB( Mat &bgr3u, Mat &mag1u );
@@ -479,7 +499,7 @@ class CV_EXPORTS_W ObjectnessBING : public Objectness
     return abs( u[0] - v[0] ) + abs( u[1] - v[1] ) + abs( u[2] - v[2] );
   }
 
-  //Non-maximal suppress
+//Non-maximal suppress
   static void nonMaxSup( Mat &matchCost1f, ValStructVec<float, Point> &matchCost, int NSS = 1, int maxPoint = 50, bool fast = true );
 
 };

modules/saliency/samples/computeSaliency.cpp

@@ -64,6 +64,7 @@ static void help()
 
 int main( int argc, char** argv )
 {
+
   CommandLineParser parser( argc, argv, keys );
 
   String saliency_algorithm = parser.get<String>( 0 );
@@ -94,13 +95,7 @@ int main( int argc, char** argv )
   Mat frame;
 
   //instantiates the specific Saliency
-  Ptr<Saliency> saliencyAlgorithm = Saliency::create( saliency_algorithm );
-
-  if( saliencyAlgorithm == NULL )
-  {
-    cout << "***Error in the instantiation of the saliency algorithm...***\n";
-    return -1;
-  }
+  Ptr<Saliency> saliencyAlgorithm;
 
   Mat binaryMap;
   Mat image;
@@ -116,6 +111,7 @@ int main( int argc, char** argv )
   if( saliency_algorithm.find( "SPECTRAL_RESIDUAL" ) == 0 )
   {
     Mat saliencyMap;
+    saliencyAlgorithm = StaticSaliencySpectralResidual::create();
     if( saliencyAlgorithm->computeSaliency( image, saliencyMap ) )
     {
       StaticSaliencySpectralResidual spec;
@@ -131,6 +127,7 @@ int main( int argc, char** argv )
   else if( saliency_algorithm.find( "FINE_GRAINED" ) == 0 )
   {
     Mat saliencyMap;
+    saliencyAlgorithm = StaticSaliencyFineGrained::create();
     if( saliencyAlgorithm->computeSaliency( image, saliencyMap ) )
     {
       imshow( "Saliency Map", saliencyMap );
@@ -150,6 +147,7 @@ int main( int argc, char** argv )
 
     else
     {
+      saliencyAlgorithm = ObjectnessBING::create();
       vector<Vec4i> saliencyMap;
       saliencyAlgorithm.dynamicCast<ObjectnessBING>()->setTrainingPath( training_path );
       saliencyAlgorithm.dynamicCast<ObjectnessBING>()->setBBResDir( training_path + "/Results" );
@@ -163,8 +161,7 @@ int main( int argc, char** argv )
   }
   else if( saliency_algorithm.find( "BinWangApr2014" ) == 0 )
   {
-
-    //Ptr<Size> size = Ptr<Size>( new Size( image.cols, image.rows ) );
+    saliencyAlgorithm = MotionSaliencyBinWangApr2014::create();
     saliencyAlgorithm.dynamicCast<MotionSaliencyBinWangApr2014>()->setImagesize( image.cols, image.rows );
     saliencyAlgorithm.dynamicCast<MotionSaliencyBinWangApr2014>()->init();
 
@@ -175,13 +172,14 @@ int main( int argc, char** argv )
       {
 
         cap >> frame;
+        if( frame.empty() )
+        {
+          return 0;
+        }
         cvtColor( frame, frame, COLOR_BGR2GRAY );
 
         Mat saliencyMap;
-        if( saliencyAlgorithm->computeSaliency( frame, saliencyMap ) )
-        {
-          std::cout << "current frame motion saliency done" << std::endl;
-        }
+        saliencyAlgorithm->computeSaliency( frame, saliencyMap );
 
         imshow( "image", frame );
         imshow( "saliencyMap", saliencyMap * 255 );