removed unnecessary memory allocation in LBP classifier

diff --git a/modules/gpu/include/opencv2/gpu/gpu.hpp b/modules/gpu/include/opencv2/gpu/gpu.hpp
index f8a16ee..4c3af59 100644 (file)
--- a/modules/gpu/include/opencv2/gpu/gpu.hpp
+++ b/modules/gpu/include/opencv2/gpu/gpu.hpp
@@ -1428,7 +1428,7 @@ class CV_EXPORTS CascadeClassifier_GPU_LBP
 public:\r
     enum stage { BOOST = 0 };\r
     enum feature { LBP = 0 };\r
-    CascadeClassifier_GPU_LBP();\r
+    CascadeClassifier_GPU_LBP(cv::Size detectionFrameSize = cv::Size());\r
     ~CascadeClassifier_GPU_LBP();\r
 \r
     bool empty() const;\r
@@ -1441,6 +1441,7 @@ public:
     Size getClassifierSize() const;\r
 private:\r
     bool read(const FileNode &root);\r
+    void initializeBuffers(cv::Size frame);\r
 \r
     static const stage stageType = BOOST;\r
     static const feature featureType = LBP;\r
@@ -1459,8 +1460,9 @@ private:
     GpuMat subsets_mat;\r
     GpuMat features_mat;\r
 \r
-    // current integral image\r
     GpuMat integral;\r
+    GpuMat integralBuffer;\r
+    GpuMat resuzeBuffer;\r
 };\r
 \r
 ////////////////////////////////// SURF //////////////////////////////////////////\r

diff --git a/modules/gpu/perf/perf_objdetect.cpp b/modules/gpu/perf/perf_objdetect.cpp
index 5fd2297..e199048 100644 (file)
--- a/modules/gpu/perf/perf_objdetect.cpp
+++ b/modules/gpu/perf/perf_objdetect.cpp
@@ -66,12 +66,12 @@ GPU_PERF_TEST_1(LBPClassifier, cv::gpu::DeviceInfo)
     cv::Mat img_host = readImage("gpu/haarcascade/group_1_640x480_VGA.pgm", cv::IMREAD_GRAYSCALE);\r
     ASSERT_FALSE(img_host.empty());\r
 \r
-    cv::gpu::CascadeClassifier_GPU_LBP cascade;\r
 \r
-    ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml")));\r
 \r
     cv::gpu::GpuMat img(img_host);\r
         cv::gpu::GpuMat gpu_rects, buffer;\r
+    cv::gpu::CascadeClassifier_GPU_LBP cascade(img.size());\r
+    ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml")));\r
 \r
     // cascade.detectMultiScale(img, objects_buffer);\r
     cascade.detectMultiScale(img, buffer, gpu_rects);\r

diff --git a/modules/gpu/src/cascadeclassifier.cpp b/modules/gpu/src/cascadeclassifier.cpp
index da0bfe5..e852256 100644 (file)
--- a/modules/gpu/src/cascadeclassifier.cpp
+++ b/modules/gpu/src/cascadeclassifier.cpp
@@ -61,20 +61,46 @@ Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const  { throw_nogpu();
 int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat& , GpuMat& , double , int , Size)  { throw_nogpu(); return 0; }\r
 \r
 // ============ LBP cascade ==============================================//\r
-cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP()               { throw_nogpu(); }\r
-cv::gpu::CascadeClassifier_GPU_LBP::~CascadeClassifier_GPU_LBP()              { throw_nogpu(); }\r
+cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP(cv::Size /*frameSize*/){ throw_nogpu(); }\r
+cv::gpu::CascadeClassifier_GPU_LBP::~CascadeClassifier_GPU_LBP()                     { throw_nogpu(); }\r
 \r
 bool cv::gpu::CascadeClassifier_GPU_LBP::empty() const                               { throw_nogpu(); return true; }\r
 bool cv::gpu::CascadeClassifier_GPU_LBP::load(const string&)                         { throw_nogpu(); return true; }\r
 Size cv::gpu::CascadeClassifier_GPU_LBP::getClassifierSize() const                   { throw_nogpu(); return Size(); }\r
 void cv::gpu::CascadeClassifier_GPU_LBP::preallocateIntegralBuffer(cv::Size /*desired*/) { throw_nogpu();}\r
+void cv::gpu::CascadeClassifier_GPU_LBP::initializeBuffers(cv::Size /*frame*/)       { throw_nogpu();}\r
 \r
 int cv::gpu::CascadeClassifier_GPU_LBP::detectMultiScale(const cv::gpu::GpuMat& /*image*/, cv::gpu::GpuMat& /*scaledImageBuffer*/, cv::gpu::GpuMat& /*objectsBuf*/,\r
 double /*scaleFactor*/, int /*minNeighbors*/, cv::Size /*maxObjectSize*/){ throw_nogpu(); return 0;}\r
 \r
 #else\r
 \r
-cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP(){}\r
+cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP(cv::Size detectionFrameSize)\r
+{\r
+    if (detectionFrameSize != cv::Size())\r
+        initializeBuffers(detectionFrameSize);\r
+}\r
+\r
+void cv::gpu::CascadeClassifier_GPU_LBP::initializeBuffers(cv::Size frame)\r
+{\r
+    if (resuzeBuffer.empty() || frame.width > resuzeBuffer.cols || frame.height > resuzeBuffer.rows)\r
+    {\r
+        resuzeBuffer.create(frame, CV_8UC1);\r
+\r
+        integral.create(frame.height + 1, frame.width + 1, CV_32SC1);\r
+        NcvSize32u roiSize;\r
+        roiSize.width = frame.width;\r
+        roiSize.height = frame.height;\r
+\r
+        cudaDeviceProp prop;\r
+        cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );\r
+\r
+        Ncv32u bufSize;\r
+        ncvSafeCall( nppiStIntegralGetSize_8u32u(roiSize, &bufSize, prop) );\r
+        // printf("HERE!!!!!!!%d\n", bufSize);\r
+        integralBuffer.create(1, bufSize, CV_8UC1);\r
+    }\r
+}\r
 \r
 cv::gpu::CascadeClassifier_GPU_LBP::~CascadeClassifier_GPU_LBP(){}\r
 \r
@@ -309,10 +335,12 @@ int cv::gpu::CascadeClassifier_GPU_LBP::detectMultiScale(const GpuMat& image, Gp
         objects.create(1 , defaultObjSearchNum, CV_32SC4);\r
 \r
     GpuMat candidates(1 , defaultObjSearchNum, CV_32SC4);\r
+    // GpuMat candidates(objects);\r
     if (maxObjectSize == cv::Size())\r
         maxObjectSize = image.size();\r
 \r
-    scaledImageBuffer.create(image.rows + 1, image.cols + 1, CV_8U);\r
+    initializeBuffers(image.size());\r
+\r
     unsigned int* classified = new unsigned int[1];\r
     *classified = 0;\r
     unsigned int* dclassified;\r
@@ -335,13 +363,17 @@ int cv::gpu::CascadeClassifier_GPU_LBP::detectMultiScale(const GpuMat& image, Gp
         // if( windowSize.width < minObjectSize.width || windowSize.height < minObjectSize.height )\r
         //     continue;\r
 \r
-        cv::gpu::resize(image, scaledImageBuffer, scaledImageSize, 0, 0, CV_INTER_LINEAR);\r
-        cv::gpu::integral(scaledImageBuffer, integral);\r
+        GpuMat scaledImg(resuzeBuffer, cv::Rect(0, 0, scaledImageSize.width, scaledImageSize.height));\r
+        GpuMat scaledIntegral(integral, cv::Rect(0, 0, scaledImageSize.width + 1, scaledImageSize.height + 1));\r
+        GpuMat currBuff = integralBuffer;//(integralBuffer, cv::Rect(0, 0, integralBuffer.width, integralBuffer.height));\r
+\r
+        cv::gpu::resize(image, scaledImg, scaledImageSize, 0, 0, CV_INTER_LINEAR);\r
+        cv::gpu::integralBuffered(scaledImg, scaledIntegral, currBuff);\r
 \r
         step = (factor <= 2.) + 1;\r
 \r
         cv::gpu::device::lbp::classifyStump(stage_mat, stage_mat.cols / sizeof(Stage), nodes_mat, leaves_mat, subsets_mat, features_mat,\r
-        integral, processingRectSize.width, processingRectSize.height, windowSize.width, windowSize.height, factor, step, subsetSize, candidates, dclassified);\r
+        scaledIntegral, processingRectSize.width, processingRectSize.height, windowSize.width, windowSize.height, factor, step, subsetSize, candidates, dclassified);\r
     }\r
     if (groupThreshold <= 0  || objects.empty())\r
         return 0;\r