Fixing problems of compilation in the GPU

CMakeLists.txt

@@ -11,7 +11,11 @@ if (CUDA_FOUND)
 endif (CUDA_FOUND)
 
 if (WITH_CUDA)
-  include_directories(${CUDA_INCLUDE_DIRS})
+  include_directories(
+    ${CUDA_INCLUDE_DIRS}
+    include
+    )
+  add_definitions(-DWITH_CUDA)
   set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -std=c++11 -arch=sm_30" )
   cuda_add_library(cu_ball_track SHARED
     src/cuda/img_proc.cu
@@ -22,7 +26,6 @@ if (WITH_CUDA)
   set(GPU_BT_LIB
     cu_ball_track
     )
-  add_definitions(-DWITH_CUDA)
 endif(WITH_CUDA)
 
 include_directories(include
@@ -40,6 +43,7 @@ target_link_libraries(ball_tracking
   ${GPU_BT_LIB}
   )
 
+
 #Compile with C++11 support only
 if(NOT MSVC AND (("${CMAKE_CXX_COMPILER_ID}" MATCHES "GNU") OR ("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")))
   include(CheckCXXCompilerFlag)
@@ -66,6 +70,3 @@ install(TARGETS ball_tracking DESTINATION lib)
 install(FILES include/json.hpp DESTINATION include)
 install(DIRECTORY include/ball_tracking DESTINATION include)
 
-if (PYLIB)
-
-endif (PYLIB)

examples/tracking/gpu_track.cpp

@@ -123,7 +123,8 @@ int main(int argc, char** argv) {
 
       //3) Run blob detection algorithm (In CPU)
       start_t = std::chrono::steady_clock::now();
-      auto key_pts = bf(bin_img);
+      gpu_l_lh.download(l_lh);
+      auto key_pts = bf(l_lh);
       end_t = std::chrono::steady_clock::now();
       blob_time.push_back(std::chrono::duration_cast<std::chrono::microseconds>(end_t - start_t).count());
     
@@ -133,7 +134,6 @@ int main(int argc, char** argv) {
 
       //4) Compute output image
       start_t = std::chrono::steady_clock::now();
-      gpu_l_lh.download(l_lh);
       drawKeypoints(img, key_pts, res(Rect(0,0,in_size.width,in_size.height)),
           Scalar(255,255,0), DrawMatchesFlags::DRAW_RICH_KEYPOINTS);
       normalize(l_lh, tmp, 0, 1, NORM_MINMAX);

src/tracker.cpp

@@ -5,6 +5,11 @@
 #include <algorithm>
 #include <json.hpp>
 
+#ifdef WITH_CUDA
+#include <ball_tracking/cuda/tracker.hpp>
+#include <opencv2/core.hpp>
+#endif
+
 using namespace cv;
 using namespace std;
 using json = nlohmann::json;
@@ -235,8 +240,9 @@ namespace ball_tracking {
     };
 
 #ifdef WITH_CUDA
-#include <ball_tracking/cuda/tracker.hpp>
-#include <opencv2/core.hpp>
+
+    using namespace cv::cuda;
+
     class GPUTracker {
       private:
         ball_tracking::cuda::BallLogLikelihood llh; //!< Fully implemented in GPU
@@ -250,14 +256,14 @@ namespace ball_tracking {
 
         vector<KeyPoint> operator()(cv::InputArray _img) {
           Mat img = _img.getMat();
-          GpuMat gpu_img(img, stream);
+          GpuMat gpu_img;
           GpuMat gpu_llh_img;
+          gpu_img.upload(img, stream);
           llh(gpu_img, gpu_llh_img, stream);
 
           Mat llh_img;
           gpu_llh_img.download(llh_img, stream);
           stream.waitForCompletion();
-
           return blob_detect(llh_img);
         }
     };