integrate speprocessing strategy

modules/gpu/include/opencv2/gpu/gpu.hpp

@@ -1529,32 +1529,37 @@ public:
 
 // ======================== GPU version for soft cascade ===================== //
 
-// Implementation of soft (stageless) cascaded detector.
-class CV_EXPORTS SCascade : public Algorithm
+class CV_EXPORTS ChannelsProcessor
 {
 public:
-
-    enum { GENERIC = 1, SEPARABLE = 2};
-    class CV_EXPORTS Preprocessor
+    enum
     {
-    public:
+        GENERIC   = 1 << 4,
+        SEPARABLE = 2 << 4
+    };
 
     // Appends specified number of HOG first-order features integrals into given vector.
     // Param frame is an input 3-channel bgr image.
-        // Param channels is a GPU matrix of integrals.
+    // Param channels is a GPU matrix of optionally shrinked channels
     // Param stream is stream is a high-level CUDA stream abstraction used for asynchronous execution.
     virtual void apply(InputArray frame, OutputArray channels, Stream& stream = Stream::Null()) = 0;
 
     // Creates a specific preprocessor implementation.
     // Param shrinkage is a resizing factor. Resize is applied before the computing integral sum
     // Param bins is a number of HOG-like channels.
-        // Param method is a channel computing method.
-        static cv::Ptr<Preprocessor> create(const int shrinkage, const int bins, const int method = GENERIC);
+    // Param flags is a channel computing extra flags.
+    static cv::Ptr<ChannelsProcessor> create(const int shrinkage, const int bins, const int flags = GENERIC);
 
+    virtual ~ChannelsProcessor();
 
-    protected:
-        Preprocessor();
-    };
+protected:
+    ChannelsProcessor();
+};
+
+// Implementation of soft (stageless) cascaded detector.
+class CV_EXPORTS SCascade : public Algorithm
+{
+public:
 
     // Representation of detectors result.
     struct CV_EXPORTS Detection
@@ -1569,14 +1574,15 @@ public:
         enum {PEDESTRIAN = 0};
     };
 
-    enum { NO_REJECT = 1, DOLLAR = 2, /*PASCAL = 4,*/ DEFAULT = NO_REJECT};
+    enum { NO_REJECT = 1, DOLLAR = 2, /*PASCAL = 4,*/ DEFAULT = NO_REJECT, NMS_MASK = 0xF};
 
     // An empty cascade will be created.
     // Param minScale is a minimum scale relative to the original size of the image on which cascade will be applyed.
     // Param minScale is a maximum scale relative to the original size of the image on which cascade will be applyed.
     // Param scales is a number of scales from minScale to maxScale.
-    // Param rejfactor is used for NMS.
-    SCascade(const double minScale = 0.4, const double maxScale = 5., const int scales = 55, const int rejCriteria = 1);
+    // Param flags is an extra tuning flags.
+    SCascade(const double minScale = 0.4, const double maxScale = 5., const int scales = 55,
+        const int flags = NO_REJECT || ChannelsProcessor::GENERIC);
 
     virtual ~SCascade();
 
@@ -1598,13 +1604,6 @@ public:
     // Param stream is stream is a high-level CUDA stream abstraction used for asynchronous execution
     virtual void detect(InputArray image, InputArray rois, OutputArray objects, Stream& stream = Stream::Null()) const;
 
-    // Convert ROI matrix into the suitable for detect method.
-    // Param roi is an input matrix of the same size as the image.
-    //    There non zero value mean that detector should be executed in this point.
-    // Param mask is an output mask
-    // Param stream is stream is a high-level CUDA stream abstraction used for asynchronous execution
-    virtual void genRoi(InputArray roi, OutputArray mask, Stream& stream = Stream::Null()) const;
-
 private:
 
     struct Fields;
@@ -1612,9 +1611,9 @@ private:
 
     double minScale;
     double maxScale;
-
     int scales;
-    int rejCriteria;
+
+    int flags;
 };
 
 CV_EXPORTS bool initModule_gpu(void);

modules/gpu/perf/perf_softcascade.cpp

@@ -71,15 +71,14 @@ RUN_GPU(SCascadeTest, detect)
 
     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 
-    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1), trois;
+    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
     rois.setTo(1);
-    cascade.genRoi(rois, trois);
 
-    cascade.detect(colored, trois, objectBoxes);
+    cascade.detect(colored, rois, objectBoxes);
 
     TEST_CYCLE()
     {
-        cascade.detect(colored, trois, objectBoxes);
+        cascade.detect(colored, rois, objectBoxes);
     }
 
     SANITY_CHECK(sortDetections(objectBoxes));
@@ -142,14 +141,11 @@ RUN_GPU(SCascadeTestRoi, detectInRoi)
         sub.setTo(1);
     }
 
-    cv::gpu::GpuMat trois;
-    cascade.genRoi(rois, trois);
-
-    cascade.detect(colored, trois, objectBoxes);
+    cascade.detect(colored, rois, objectBoxes);
 
     TEST_CYCLE()
     {
-        cascade.detect(colored, trois, objectBoxes);
+        cascade.detect(colored, rois, objectBoxes);
     }
 
     SANITY_CHECK(sortDetections(objectBoxes));
@@ -186,14 +182,11 @@ RUN_GPU(SCascadeTestRoi, detectEachRoi)
     cv::gpu::GpuMat sub(rois, r);
     sub.setTo(1);
 
-    cv::gpu::GpuMat trois;
-    cascade.genRoi(rois, trois);
-
-    cascade.detect(colored, trois, objectBoxes);
+    cascade.detect(colored, rois, objectBoxes);
 
     TEST_CYCLE()
     {
-        cascade.detect(colored, trois, objectBoxes);
+        cascade.detect(colored, rois, objectBoxes);
     }
 
     SANITY_CHECK(sortDetections(objectBoxes));
@@ -235,15 +228,14 @@ RUN_GPU(SCascadeTest, detectOnIntegral)
 
     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 
-    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(cv::Size(640, 480), CV_8UC1), trois;
+    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(cv::Size(640, 480), CV_8UC1);
     rois.setTo(1);
-    cascade.genRoi(rois, trois);
 
-    cascade.detect(hogluv, trois, objectBoxes);
+    cascade.detect(hogluv, rois, objectBoxes);
 
     TEST_CYCLE()
     {
-        cascade.detect(hogluv, trois, objectBoxes);
+        cascade.detect(hogluv, rois, objectBoxes);
     }
 
     SANITY_CHECK(sortDetections(objectBoxes));
@@ -270,18 +262,16 @@ RUN_GPU(SCascadeTest, detectStream)
 
     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 
-    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1), trois;
+    cv::gpu::GpuMat objectBoxes(1, 10000 * sizeof(cv::gpu::SCascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
     rois.setTo(1);
 
     cv::gpu::Stream s;
 
-    cascade.genRoi(rois, trois, s);
-
-    cascade.detect(colored, trois, objectBoxes, s);
+    cascade.detect(colored, rois, objectBoxes, s);
 
     TEST_CYCLE()
     {
-        cascade.detect(colored, trois, objectBoxes, s);
+        cascade.detect(colored, rois, objectBoxes, s);
     }
 
 #ifdef HAVE_CUDA

modules/gpu/src/gpu_init.cpp

@@ -48,8 +48,7 @@ namespace cv { namespace gpu
 CV_INIT_ALGORITHM(SCascade, "CascadeDetector.SCascade",
                   obj.info()->addParam(obj, "minScale", obj.minScale);
                   obj.info()->addParam(obj, "maxScale", obj.maxScale);
-                  obj.info()->addParam(obj, "scales",      obj.scales);
-                  obj.info()->addParam(obj, "rejCriteria", obj.rejCriteria));
+                  obj.info()->addParam(obj, "scales",   obj.scales));
 
 bool initModule_gpu(void)
 {

modules/gpu/test/test_softcascade.cpp

@@ -169,7 +169,7 @@ GPU_TEST_P(SCascadeTestRoi, detect,
 
     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 
-    GpuMat colored(coloredCpu), objectBoxes(1, 16384, CV_8UC1), rois(colored.size(), CV_8UC1), trois;
+    GpuMat colored(coloredCpu), objectBoxes(1, 16384, CV_8UC1), rois(colored.size(), CV_8UC1);
     rois.setTo(0);
 
     int nroi = GET_PARAM(3);
@@ -183,8 +183,8 @@ GPU_TEST_P(SCascadeTestRoi, detect,
         cv::rectangle(result, r, cv::Scalar(0, 0, 255, 255), 1);
     }
     objectBoxes.setTo(0);
-    cascade.genRoi(rois, trois);
-    cascade.detect(colored, trois, objectBoxes);
+
+    cascade.detect(colored, rois, objectBoxes);
 
     cv::Mat dt(objectBoxes);
     typedef cv::gpu::SCascade::Detection Detection;
@@ -239,10 +239,8 @@ GPU_TEST_P(SCascadeTestAll, detect,
     GpuMat sub(rois, cv::Rect(rois.cols / 4, rois.rows / 4,rois.cols / 2, rois.rows / 2));
     sub.setTo(cv::Scalar::all(1));
 
-    cv::gpu::GpuMat trois;
-    cascade.genRoi(rois, trois);
     objectBoxes.setTo(0);
-    cascade.detect(colored, trois, objectBoxes);
+    cascade.detect(colored, rois, objectBoxes);
 
     typedef cv::gpu::SCascade::Detection Detection;
     cv::Mat detections(objectBoxes);
@@ -279,10 +277,8 @@ GPU_TEST_P(SCascadeTestAll, detectOnIntegral,
     GpuMat objectBoxes(1, 100000, CV_8UC1), rois(cv::Size(640, 480), CV_8UC1);
     rois.setTo(1);
 
-    cv::gpu::GpuMat trois;
-    cascade.genRoi(rois, trois);
     objectBoxes.setTo(0);
-    cascade.detect(hogluv, trois, objectBoxes);
+    cascade.detect(hogluv, rois, objectBoxes);
 
     typedef cv::gpu::SCascade::Detection Detection;
     cv::Mat detections(objectBoxes);
@@ -315,12 +311,9 @@ GPU_TEST_P(SCascadeTestAll, detectStream,
 
     cv::gpu::Stream s;
 
-    cv::gpu::GpuMat trois;
-    cascade.genRoi(rois, trois, s);
     objectBoxes.setTo(0);
-    cascade.detect(colored, trois, objectBoxes, s);
-
-    cudaDeviceSynchronize();
+    cascade.detect(colored, rois, objectBoxes, s);
+    s.waitForCompletion();
 
     typedef cv::gpu::SCascade::Detection Detection;
     cv::Mat detections(objectBoxes);