Moving to AVX instead of AVX2, compiler warning fixes. Losing dependency on OpenCV C style bindings for affine warping.

lib/local/CppInerop/CppInerop.vcxproj

@@ -145,7 +145,7 @@
       <PreprocessorDefinitions>WIN32;NDEBUG;_WINDOWS;_USRDLL;CPPINEROP_EXPORTS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>./;$(SolutionDir)lib\local\LandmarkDetector\include;$(SolutionDir)lib\local\FaceAnalyser\include;$(SolutionDir)lib\local\GazeAnalyser\include;$(SolutionDir)lib\local\Utilities\include;$(SolutionDir)lib\3rdParty\CameraEnumerator;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
     </ClCompile>
     <Link>
       <SubSystem>Windows</SubSystem>
@@ -166,7 +166,7 @@
       <PreprocessorDefinitions>NDEBUG;_WINDOWS;_USRDLL;CPPINEROP_EXPORTS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>./;$(SolutionDir)lib\local\LandmarkDetector\include;$(SolutionDir)lib\local\FaceAnalyser\include;$(SolutionDir)lib\local\GazeAnalyser\include;$(SolutionDir)lib\local\Utilities\include;$(SolutionDir)lib\3rdParty\CameraEnumerator;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
     </ClCompile>
     <Link>
       <SubSystem>Windows</SubSystem>

lib/local/FaceAnalyser/AU_predictors/AU_DISFA_best.txt

@@ -0,0 +1,24 @@
+svm_disfa/AU_1_dynamic.dat AU01
+svm_disfa/AU_2_dynamic.dat AU02
+svm_disfa/AU_4_static.dat AU04
+svm_disfa/AU_5_dynamic.dat AU05
+svm_disfa/AU_6_static.dat AU06
+svm_disfa/AU_9_dynamic.dat AU09
+svm_disfa/AU_12_dynamic.dat AU12
+svm_disfa/AU_15_dynamic.dat AU15
+svm_disfa/AU_17_dynamic.dat AU17
+svm_disfa/AU_20_dynamic.dat AU20
+svm_disfa/AU_25_dynamic.dat AU25
+svm_disfa/AU_26_dynamic.dat AU26
+svr_disfa/AU_1_dynamic_intensity.dat AU01
+svr_disfa/AU_2_dynamic_intensity.dat AU02
+svr_disfa/AU_4_dynamic_intensity_comb.dat AU04
+svr_disfa/AU_5_dynamic_intensity.dat AU05
+svr_disfa/AU_6_dynamic_intensity_comb.dat AU06
+svr_disfa/AU_9_dynamic_intensity.dat AU09
+svr_disfa/AU_12_dynamic_intensity_comb.dat AU12
+svr_disfa/AU_15_dynamic_intensity_comb.dat AU15
+svr_disfa/AU_17_dynamic_intensity_comb.dat AU17
+svr_disfa/AU_20_dynamic_intensity.dat AU20
+svr_disfa/AU_25_dynamic_intensity_comb.dat AU25
+svr_disfa/AU_26_dynamic_intensity_comb.dat AU26

lib/local/FaceAnalyser/AU_predictors/AU_DISFA_static.txt

@@ -0,0 +1,24 @@
+svm_disfa/AU_1_static.dat AU01
+svm_disfa/AU_2_static.dat AU02
+svm_disfa/AU_4_static.dat AU04
+svm_disfa/AU_5_static.dat AU05
+svm_disfa/AU_6_static.dat AU06
+svm_disfa/AU_9_static.dat AU09
+svm_disfa/AU_12_static.dat AU12
+svm_disfa/AU_15_static.dat AU15
+svm_disfa/AU_17_static.dat AU17
+svm_disfa/AU_20_static.dat AU20
+svm_disfa/AU_25_static.dat AU25
+svm_disfa/AU_26_static.dat AU26
+svr_disfa/AU_1_static_intensity.dat AU01
+svr_disfa/AU_2_static_intensity.dat AU02
+svr_disfa/AU_4_static_intensity.dat AU04
+svr_disfa/AU_5_static_intensity.dat AU05
+svr_disfa/AU_6_static_intensity.dat AU06
+svr_disfa/AU_9_static_intensity.dat AU09
+svr_disfa/AU_12_static_intensity.dat AU12
+svr_disfa/AU_15_static_intensity.dat AU15
+svr_disfa/AU_17_static_intensity.dat AU17
+svr_disfa/AU_20_static_intensity.dat AU20
+svr_disfa/AU_25_static_intensity.dat AU25
+svr_disfa/AU_26_static_intensity.dat AU26

lib/local/FaceAnalyser/AU_predictors/DISFA_dynamic_svms.txt

@@ -0,0 +1,3 @@
+AUPredictor AU_DISFA_best.txt
+PDM In-the-wild_aligned_PDM_68.txt
+Triangulation tris_68_full.txt

lib/local/FaceAnalyser/AU_predictors/DISFA_static_svms.txt

@@ -0,0 +1,3 @@
+AUPredictor AU_DISFA_static.txt
+PDM In-the-wild_aligned_PDM_68.txt
+Triangulation tris_68_full.txt

lib/local/FaceAnalyser/FaceAnalyser.vcxproj

@@ -141,7 +141,7 @@
       <SDLCheck>
       </SDLCheck>
       <AdditionalIncludeDirectories>./include;$(SolutionDir)lib/local/Utilities/include;$(SolutionDir)lib/local/LandmarkDetector/include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
     </ClCompile>
@@ -167,7 +167,7 @@
       <SDLCheck>
       </SDLCheck>
       <AdditionalIncludeDirectories>./include;$(SolutionDir)lib/local/Utilities/include;$(SolutionDir)lib/local/LandmarkDetector/include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <PreprocessorDefinitions>WIN64;NDEBUG;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>

lib/local/GazeAnalyser/GazeAnalyser.vcxproj

@@ -116,7 +116,7 @@
       <AdditionalIncludeDirectories>./include;../LandmarkDetector/include;../Utilities/include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <PreprocessorDefinitions>WIN32;NDEBUG;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
     </ClCompile>
     <Link>
@@ -135,7 +135,7 @@
       <AdditionalIncludeDirectories>./include;../LandmarkDetector/include;../Utilities/include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <PreprocessorDefinitions>WIN64;NDEBUG;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
     </ClCompile>
     <Link>

lib/local/LandmarkDetector/LandmarkDetector.vcxproj

@@ -148,7 +148,7 @@ xcopy /I /E /Y /D "$(SolutionDir)lib\3rdParty\OpenCV3.4\classifiers" "$(OutDir)c
       <PrecompiledHeader>Use</PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
       <OpenMPSupport>false</OpenMPSupport>
       <AdditionalOptions>/Zm300 %(AdditionalOptions)</AdditionalOptions>
@@ -173,7 +173,7 @@ xcopy /I /E /Y /D "$(SolutionDir)lib\3rdParty\OpenCV3.4\classifiers" "$(OutDir)c
       <PrecompiledHeader>Use</PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
       <OpenMPSupport>false</OpenMPSupport>
       <AdditionalOptions>/Zm300 %(AdditionalOptions)</AdditionalOptions>

lib/local/LandmarkDetector/include/CEN_patch_expert.h

@@ -53,8 +53,8 @@ namespace LandmarkDetector
 	public:
 
 		// Width and height of the patch expert support area
-		int width;
-		int height;
+		int width_support;
+		int height_support;
 
 		// Neural weights
 		std::vector<cv::Mat_<float>> biases;

lib/local/LandmarkDetector/src/CCNF_patch_expert.cpp

@@ -455,8 +455,7 @@ void CCNF_patch_expert::ResponseOpenBlas(const cv::Mat_<float> &area_of_interest
 	// Perform matrix multiplication in OpenBLAS (fortran call)
 	float alpha1 = 1.0;
 	float beta1 = 0.0;
-	char *nT = "N";
-	sgemm_(nT, nT, &normalized_input.cols, &weight_matrix.rows, &weight_matrix.cols, &alpha1, (float*)normalized_input.data, &normalized_input.cols, (float*)weight_matrix.data, &weight_matrix.cols, &beta1, (float*)neuron_resp_full.data, &normalized_input.cols);
+	sgemm_("N", "N", &normalized_input.cols, &weight_matrix.rows, &weight_matrix.cols, &alpha1, (float*)normalized_input.data, &normalized_input.cols, (float*)weight_matrix.data, &weight_matrix.cols, &beta1, (float*)neuron_resp_full.data, &normalized_input.cols);
 
 	// Above is a faster version of this
 	//cv::Mat_<float> neuron_resp_full = this->weight_matrix * normalized_input;
@@ -500,8 +499,7 @@ void CCNF_patch_expert::ResponseOpenBlas(const cv::Mat_<float> &area_of_interest
 	// Perform matrix multiplication in OpenBLAS (fortran call)
 	alpha1 = 1.0;
 	beta1 = 0.0;
-	nT = "N";
-	sgemm_(nT, nT, &resp_vec_f.cols, &Sigmas[s_to_use].rows, &Sigmas[s_to_use].cols, &alpha1, (float*)resp_vec_f.data, &resp_vec_f.cols, (float*)Sigmas[s_to_use].data, &Sigmas[s_to_use].cols, &beta1, (float*)out.data, &resp_vec_f.cols);
+	sgemm_("N", "N", &resp_vec_f.cols, &Sigmas[s_to_use].rows, &Sigmas[s_to_use].cols, &alpha1, (float*)resp_vec_f.data, &resp_vec_f.cols, (float*)Sigmas[s_to_use].data, &Sigmas[s_to_use].cols, &beta1, (float*)out.data, &resp_vec_f.cols);
 
 	// Above is a faster version of this
 	//cv::Mat out = Sigmas[s_to_use] * resp_vec_f;

lib/local/LandmarkDetector/src/CEN_patch_expert.cpp

@@ -77,7 +77,7 @@
 using namespace LandmarkDetector;
 
 // Copy constructor		
-CEN_patch_expert::CEN_patch_expert(const CEN_patch_expert& other) : confidence(other.confidence), width(other.width), height(other.height)
+CEN_patch_expert::CEN_patch_expert(const CEN_patch_expert& other) : confidence(other.confidence), width_support(other.width_support), height_support(other.height_support)
 {
 
 	// Copy the layer weights in a deep way
@@ -106,8 +106,8 @@ void CEN_patch_expert::Read(ifstream &stream)
 
 	// the number of neurons for this patch
 	int num_layers;
-	stream.read((char*)&width, 4);
-	stream.read((char*)&height, 4);
+	stream.read((char*)&width_support, 4);
+	stream.read((char*)&height_support, 4);
 	stream.read((char*)&num_layers, 4);
 
 	if (num_layers == 0)
@@ -221,11 +221,11 @@ void im2colBias(const cv::Mat_<float>& input, int width, int height, cv::Mat_<fl
 void CEN_patch_expert::Response(const cv::Mat_<float> &area_of_interest, cv::Mat_<float> &response)
 {
 
-	int response_height = area_of_interest.rows - height + 1;
-	int response_width = area_of_interest.cols - width + 1;
+	int response_height = area_of_interest.rows - height_support + 1;
+	int response_width = area_of_interest.cols - width_support + 1;
 	
 	cv::Mat_<float> input_col;
-	im2colBias(area_of_interest, width, height, input_col);
+	im2colBias(area_of_interest, width_support, height_support, input_col);
 
 	// Mean and standard deviation normalization
 	contrastNorm(input_col, response);
@@ -246,8 +246,7 @@ void CEN_patch_expert::Response(const cv::Mat_<float> &area_of_interest, cv::Mat
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 0.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
+		sgemm_("N", "N", &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
 
 		// The above is a faster version of this, by calling the fortran version directly
 		//cblas_sgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, resp.cols, weight.rows, weight.cols, 1, m1, resp.cols, m2, weight.cols, 0.0, m3, resp.cols);
@@ -507,11 +506,11 @@ void LandmarkDetector::interpolationMatrix(cv::Mat_<float>& mapMatrix, int respo
 void CEN_patch_expert::ResponseSparse(const cv::Mat_<float> &area_of_interest, cv::Mat_<float> &response, cv::Mat_<float>& mapMatrix, cv::Mat_<float>& im2col_prealloc)
 {
 
-	int response_height = area_of_interest.rows - height + 1;
-	int response_width = area_of_interest.cols - width + 1;
+	int response_height = area_of_interest.rows - height_support + 1;
+	int response_width = area_of_interest.cols - width_support + 1;
 
 	// Extract im2col but in a sparse way and contrast normalize
-	im2colBiasSparseContrastNorm(area_of_interest, width, height, im2col_prealloc);
+	im2colBiasSparseContrastNorm(area_of_interest, width_support, height_support, im2col_prealloc);
 	response = im2col_prealloc.t();
 
 	for (size_t layer = 0; layer < activation_function.size(); ++layer)
@@ -529,8 +528,7 @@ void CEN_patch_expert::ResponseSparse(const cv::Mat_<float> &area_of_interest, c
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 0.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
+		sgemm_("N", "N", &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
 
 		// The above is a faster version of this, by calling the fortran version directly
 		//cblas_sgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, resp.cols, weight.rows, weight.cols, 1, m1, resp.cols, m2, weight.cols, 0.0, m3, resp.cols);
@@ -585,13 +583,13 @@ void CEN_patch_expert::ResponseSparse(const cv::Mat_<float> &area_of_interest, c
 void CEN_patch_expert::ResponseSparse_mirror(const cv::Mat_<float> &area_of_interest, cv::Mat_<float> &response, cv::Mat_<float>& mapMatrix, cv::Mat_<float>& im2col_prealloc)
 {
 
-	int response_height = area_of_interest.rows - height + 1;
-	int response_width = area_of_interest.cols - width + 1;
+	int response_height = area_of_interest.rows - height_support + 1;
+	int response_width = area_of_interest.cols - width_support + 1;
 
 	cv::flip(area_of_interest, area_of_interest, 1);
 
 	// Extract im2col but in a sparse way and contrast normalize
-	im2colBiasSparseContrastNorm(area_of_interest, width, height, im2col_prealloc);
+	im2colBiasSparseContrastNorm(area_of_interest, width_support, height_support, im2col_prealloc);
 
 	response = im2col_prealloc.t();
 
@@ -610,8 +608,7 @@ void CEN_patch_expert::ResponseSparse_mirror(const cv::Mat_<float> &area_of_inte
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 0.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
+		sgemm_("N", "N", &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
 
 		// The above is a faster version of this, by calling the fortran version directly
 		//cblas_sgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, resp.cols, weight.rows, weight.cols, 1, m1, resp.cols, m2, weight.cols, 0.0, m3, resp.cols);
@@ -665,14 +662,14 @@ void CEN_patch_expert::ResponseSparse_mirror(const cv::Mat_<float> &area_of_inte
 
 void CEN_patch_expert::ResponseSparse_mirror_joint(const cv::Mat_<float> &area_of_interest_left, const cv::Mat_<float> &area_of_interest_right, cv::Mat_<float> &response_left, cv::Mat_<float> &response_right, cv::Mat_<float>& mapMatrix, cv::Mat_<float>& im2col_prealloc_left, cv::Mat_<float>& im2col_prealloc_right)
 {
-	int response_height = area_of_interest_left.rows - height + 1;
-	int response_width = area_of_interest_left.cols - width + 1;
+	int response_height = area_of_interest_left.rows - height_support + 1;
+	int response_width = area_of_interest_left.cols - width_support + 1;
 
 	cv::flip(area_of_interest_right, area_of_interest_right, 1);
 
 	// Extract im2col but in a sparse way and contrast normalize
-	im2colBiasSparseContrastNorm(area_of_interest_left, width, height, im2col_prealloc_left);
-	im2colBiasSparseContrastNorm(area_of_interest_right, width, height, im2col_prealloc_right);
+	im2colBiasSparseContrastNorm(area_of_interest_left, width_support, height_support, im2col_prealloc_left);
+	im2colBiasSparseContrastNorm(area_of_interest_right, width_support, height_support, im2col_prealloc_right);
 
 	cv::Mat_<float> response;
 	cv::vconcat(im2col_prealloc_left, im2col_prealloc_right, response);
@@ -695,8 +692,7 @@ void CEN_patch_expert::ResponseSparse_mirror_joint(const cv::Mat_<float> &area_o
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 0.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
+		sgemm_("N", "N", &resp.cols, &weight.rows, &weight.cols, &alpha1, m1, &resp.cols, m2, &weight.cols, &beta1, m3, &resp.cols);
 
 		// The above is a faster version of this, by calling the fortran version directly
 		//cblas_sgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, resp.cols, weight.rows, weight.cols, 1, m1, resp.cols, m2, weight.cols, 0.0, m3, resp.cols);

lib/local/LandmarkDetector/src/CNN_utils.cpp

@@ -563,8 +563,7 @@ namespace LandmarkDetector
 		float alpha = 1.0f;
 		float beta = 0.0f;
 		// Call fortran directly (faster)
-		char *nT = "N";
-		sgemm_(nT, nT, &m2_cols, &num_rows, &pre_alloc_im2col.cols, &alpha, m2, &m2_cols, m1, &pre_alloc_im2col.cols, &beta, m3, &m2_cols);
+		sgemm_("N", "N", &m2_cols, &num_rows, &pre_alloc_im2col.cols, &alpha, m2, &m2_cols, m1, &pre_alloc_im2col.cols, &beta, m3, &m2_cols);
 
 		// Above is equivalent to out = pre_alloc_im2col * weight_matrix;
 		

lib/local/LandmarkDetector/src/FaceDetectorMTCNN.cpp

@@ -762,7 +762,8 @@ bool FaceDetectorMTCNN::DetectFaces(vector<cv::Rect_<float> >& o_regions, const
 	rectify(proposal_boxes_all);
 
 	// Creating proposal images from previous step detections
-	vector<bool> above_thresh(proposal_boxes_all.size());
+	vector<bool> above_thresh;
+	above_thresh.resize(proposal_boxes_all.size(), false);
 	//tbb::parallel_for(0, (int)proposal_boxes_all.size(), [&](int k) {
 	for (size_t k = 0; k < proposal_boxes_all.size(); ++k) 
 	{

lib/local/LandmarkDetector/src/LandmarkDetectorModel.cpp

@@ -1115,8 +1115,7 @@ float CLNF::NU_RLMS(cv::Vec6f& final_global, cv::Mat_<float>& final_local, const
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 1.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &J.cols, &J_w_t.rows, &J_w_t.cols, &alpha1, (float*)J.data, &J.cols, (float*)J_w_t.data, &J_w_t.cols, &beta1, (float*)Hessian.data, &J.cols);
+		sgemm_("N", "N", &J.cols, &J_w_t.rows, &J_w_t.cols, &alpha1, (float*)J.data, &J.cols, (float*)J_w_t.data, &J_w_t.cols, &beta1, (float*)Hessian.data, &J.cols);
 
 		// Above is a fast (but ugly) version of 
 		// cv::Mat_<float> Hessian = J_w_t * J + regTerm;

lib/local/LandmarkDetector/src/LandmarkDetectorUtils.cpp

@@ -429,8 +429,7 @@ namespace LandmarkDetector
 		// Get the rotation
 		cv::Matx22f R = AlignShapesKabsch2D_f(src_mean_normed, dst_mean_normed);
 
-		cv::Matx22f	A;
-		cv::Mat(s * R).copyTo(A);
+		cv::Matx22f	A = s * R;
 
 		cv::Mat_<float> aligned = (cv::Mat(cv::Mat(A) * src.t())).t();
 		cv::Mat_<float> offset = dst - aligned;

lib/local/LandmarkDetector/src/PDM.cpp

@@ -150,8 +150,7 @@ void PDM::CalcShape3D(cv::Mat_<float>& out_shape, const cv::Mat_<float>& p_local
 	int p_local_cols = p_local.cols;
 	int princ_comp_rows = princ_comp.rows;
 	int princ_comp_cols = princ_comp.cols;
-	char *nT = "N";
-	sgemm_(nT, nT, &p_local_cols, &princ_comp_rows, &princ_comp_cols, &alpha1, (float*)p_local.data, &p_local_cols, (float*)princ_comp.data, &princ_comp_cols, &beta1, (float*)out_shape.data, &p_local_cols);
+	sgemm_("N", "N", &p_local_cols, &princ_comp_rows, &princ_comp_cols, &alpha1, (float*)p_local.data, &p_local_cols, (float*)princ_comp.data, &princ_comp_cols, &beta1, (float*)out_shape.data, &p_local_cols);
 
 	// Above is a fast (but ugly) version of 
 	// out_shape = mean_shape + princ_comp * p_local;	 
@@ -639,8 +638,7 @@ void PDM::CalcParams(cv::Vec6f& out_params_global, cv::Mat_<float>& out_params_l
 		// Perform matrix multiplication in OpenBLAS (fortran call)
 		float alpha1 = 1.0;
 		float beta1 = 1.0;
-		char *nT = "N";
-		sgemm_(nT, nT, &J.cols, &J_w_t.rows, &J_w_t.cols, &alpha1, (float*)J.data, &J.cols, (float*)J_w_t.data, &J_w_t.cols, &beta1, (float*)Hessian.data, &J.cols);
+		sgemm_("N", "N", &J.cols, &J_w_t.rows, &J_w_t.cols, &alpha1, (float*)J.data, &J.cols, (float*)J_w_t.data, &J_w_t.cols, &beta1, (float*)Hessian.data, &J.cols);
 
 		// Above is a fast (but ugly) version of 
 		// cv::Mat_<float> Hessian2 = J_w_t * J + regularisations;

lib/local/LandmarkDetector/src/Patch_experts.cpp

@@ -162,10 +162,13 @@ void Patch_experts::Response(vector<cv::Mat_<float> >& patch_expert_responses, c
 
 	sim_img_to_ref = AlignShapesWithScale_f(image_shape_2D, reference_shape_2D);
 	sim_ref_to_img = sim_img_to_ref.inv(cv::DECOMP_LU);
-
+	
 	float a1 = sim_ref_to_img(0, 0);
 	float b1 = -sim_ref_to_img(0, 1);
 
+	cv::Mat_<float> gray_image_flt;
+	grayscale_image.convertTo(gray_image_flt, CV_32F);
+
 	bool use_ccnf = !this->ccnf_expert_intensity.empty();
 	bool use_cen = !this->cen_expert_intensity.empty();
 
@@ -230,8 +233,8 @@ void Patch_experts::Response(vector<cv::Mat_<float> >& patch_expert_responses, c
 
 		if (use_cen)
 		{
-			area_of_interest_width = window_size + cen_expert_intensity[scale][view_id][ind].width - 1;
-			area_of_interest_height = window_size + cen_expert_intensity[scale][view_id][ind].height - 1;
+			area_of_interest_width = window_size + cen_expert_intensity[scale][view_id][ind].width_support - 1;
+			area_of_interest_height = window_size + cen_expert_intensity[scale][view_id][ind].height_support - 1;
 		}
 		else if (use_ccnf)
 		{
@@ -245,17 +248,13 @@ void Patch_experts::Response(vector<cv::Mat_<float> >& patch_expert_responses, c
 		}
 
 		// scale and rotate to mean shape to reference frame
-		cv::Mat sim = (cv::Mat_<float>(2, 3) << a1, -b1, landmark_locations.at<float>(ind, 0), b1, a1, landmark_locations.at<float>(ind + n, 0));
+		cv::Mat sim = (cv::Mat_<float>(2, 3) << a1, -b1, landmark_locations.at<float>(ind, 0) - a1 * (area_of_interest_width - 1.0f) / 2.0f + b1 * (area_of_interest_width - 1.0f) / 2.0f, b1, a1, landmark_locations.at<float>(ind + n, 0) - a1 * (area_of_interest_width - 1.0f) / 2.0f - b1 * (area_of_interest_width - 1.0f) / 2.0f);
 
 		// Extract the region of interest around the current landmark location
-		cv::Mat_<float> area_of_interest(area_of_interest_height, area_of_interest_width);
-
-		// Using C style openCV as it does what we need
-		CvMat area_of_interest_o = area_of_interest;
-		CvMat sim_o = sim;
-		IplImage im_o = grayscale_image;
-		cvGetQuadrangleSubPix(&im_o, &area_of_interest_o, &sim_o);
+		cv::Mat_<float> area_of_interest(area_of_interest_height, area_of_interest_width, 0.0f);
 
+		cv::warpAffine(gray_image_flt, area_of_interest, sim, area_of_interest.size(), cv::WARP_INVERSE_MAP + CV_INTER_LINEAR);
+		
 
 		// Get intensity response either from the SVR, CCNF, or CEN patch experts (prefer CEN as they are the most accurate so far)
 		if (!cen_expert_intensity.empty())
@@ -281,15 +280,14 @@ void Patch_experts::Response(vector<cv::Mat_<float> >& patch_expert_responses, c
 					{
 
 						// Grab mirrored area of interest
-						cv::Mat sim_r = (cv::Mat_<float>(2, 3) << a1, -b1, landmark_locations.at<float>(mirror_id, 0), b1, a1, landmark_locations.at<float>(mirror_id + n, 0));
+
+						// scale and rotate to mean shape to reference frame
+						cv::Mat sim_r = (cv::Mat_<float>(2, 3) << a1, -b1, landmark_locations.at<float>(mirror_id, 0) - a1 * (area_of_interest_width - 1.0f) / 2.0f + b1 * (area_of_interest_width - 1.0f) / 2.0f, b1, a1, landmark_locations.at<float>(mirror_id + n, 0) - a1 * (area_of_interest_width - 1.0f) / 2.0f - b1 * (area_of_interest_width - 1.0f) / 2.0f);
 
 						// Extract the region of interest around the current landmark location
-						cv::Mat_<float> area_of_interest_r(area_of_interest_height, area_of_interest_width);
-						// Using C style openCV as it does what we need
-						CvMat area_of_interest_o_r = area_of_interest_r;
-						CvMat sim_o_r = sim_r;
-						IplImage im_o_r = grayscale_image;
-						cvGetQuadrangleSubPix(&im_o_r, &area_of_interest_o_r, &sim_o_r);
+						cv::Mat_<float> area_of_interest_r(area_of_interest_height, area_of_interest_width, 0.0f);
+
+						cv::warpAffine(gray_image_flt, area_of_interest_r, sim_r, area_of_interest_r.size(), cv::WARP_INVERSE_MAP + CV_INTER_LINEAR);
 
 						cv::Mat_<float> prealloc_mat_right = preallocated_im2col[mirror_id][im2col_size];
 

lib/local/Utilities/Utilities.vcxproj

@@ -126,7 +126,7 @@
       <AdditionalIncludeDirectories>./include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <InlineFunctionExpansion>AnySuitable</InlineFunctionExpansion>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
     </ClCompile>
     <Link>
       <SubSystem>Windows</SubSystem>
@@ -147,7 +147,7 @@
       <AdditionalIncludeDirectories>./include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
       <InlineFunctionExpansion>AnySuitable</InlineFunctionExpansion>
-      <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet>AdvancedVectorExtensions</EnableEnhancedInstructionSet>
     </ClCompile>
     <Link>
       <SubSystem>Windows</SubSystem>