Merge remote-tracking branch 'upstream/3.4' into merge-3.4

CMakeLists.txt

@@ -275,7 +275,7 @@ OCV_OPTION(WITH_VA             "Include VA support"                          OFF
 OCV_OPTION(WITH_VA_INTEL       "Include Intel VA-API/OpenCL support"         OFF  IF (UNIX AND NOT ANDROID) )
 OCV_OPTION(WITH_MFX            "Include Intel Media SDK support"             OFF   IF ((UNIX AND NOT ANDROID) OR (WIN32 AND NOT WINRT AND NOT MINGW)) )
 OCV_OPTION(WITH_GDAL           "Include GDAL Support"                        OFF  IF (NOT ANDROID AND NOT IOS AND NOT WINRT) )
-OCV_OPTION(WITH_GPHOTO2        "Include gPhoto2 library support"             ON   IF (UNIX AND NOT ANDROID AND NOT IOS) )
+OCV_OPTION(WITH_GPHOTO2        "Include gPhoto2 library support"             OFF  IF (UNIX AND NOT ANDROID AND NOT IOS) )
 OCV_OPTION(WITH_LAPACK         "Include Lapack library support"              (NOT CV_DISABLE_OPTIMIZATION)  IF (NOT ANDROID AND NOT IOS) )
 OCV_OPTION(WITH_ITT            "Include Intel ITT support"                   ON   IF (NOT APPLE_FRAMEWORK) )
 OCV_OPTION(WITH_PROTOBUF       "Enable libprotobuf"                          ON )

cmake/OpenCVDetectInferenceEngine.cmake

@@ -78,9 +78,9 @@ endif()
 
 if(INF_ENGINE_TARGET)
   if(NOT INF_ENGINE_RELEASE)
-    message(WARNING "InferenceEngine version have not been set, 2018R2 will be used by default. Set INF_ENGINE_RELEASE variable if you experience build errors.")
+    message(WARNING "InferenceEngine version have not been set, 2018R3 will be used by default. Set INF_ENGINE_RELEASE variable if you experience build errors.")
   endif()
-  set(INF_ENGINE_RELEASE "2018020000" CACHE STRING "Force IE version, should be in form YYYYAABBCC (e.g. 2018R2.0.2 -> 2018020002)")
+  set(INF_ENGINE_RELEASE "2018030000" CACHE STRING "Force IE version, should be in form YYYYAABBCC (e.g. 2018R2.0.2 -> 2018020002)")
   set_target_properties(${INF_ENGINE_TARGET} PROPERTIES
     INTERFACE_COMPILE_DEFINITIONS "HAVE_INF_ENGINE=1;INF_ENGINE_RELEASE=${INF_ENGINE_RELEASE}"
   )

cmake/OpenCVFindVA.cmake

@@ -12,7 +12,9 @@ endif()
 
 if(VA_INCLUDE_DIR)
     set(HAVE_VA TRUE)
-    set(VA_LIBRARIES "-lva" "-lva-drm")
+    if(NOT DEFINED VA_LIBRARIES)
+      set(VA_LIBRARIES "va" "va-drm")
+    endif()
 else()
     set(HAVE_VA FALSE)
     message(WARNING "libva installation is not found.")

doc/tutorials/dnn/dnn_android/dnn_android.markdown

@@ -12,7 +12,7 @@ Tutorial was written for the following versions of corresponding software:
 
 - Download and install Android Studio from https://developer.android.com/studio.
 
-- Get the latest pre-built OpenCV for Android release from https://github.com/opencv/opencv/releases and unpack it (for example, `opencv-3.4.2-android-sdk.zip`).
+- Get the latest pre-built OpenCV for Android release from https://github.com/opencv/opencv/releases and unpack it (for example, `opencv-3.4.3-android-sdk.zip`).
 
 - Download MobileNet object detection model from https://github.com/chuanqi305/MobileNet-SSD. We need a configuration file `MobileNetSSD_deploy.prototxt` and weights `MobileNetSSD_deploy.caffemodel`.
 

doc/tutorials/introduction/cross_referencing/tutorial_cross_referencing.markdown

@@ -36,14 +36,14 @@ Open your Doxyfile using your favorite text editor and search for the key
 `TAGFILES`. Change it as follows:
 
 @code
-TAGFILES = ./docs/doxygen-tags/opencv.tag=http://docs.opencv.org/3.4.2
+TAGFILES = ./docs/doxygen-tags/opencv.tag=http://docs.opencv.org/3.4.3
 @endcode
 
 If you had other definitions already, you can append the line using a `\`:
 
 @code
 TAGFILES = ./docs/doxygen-tags/libstdc++.tag=https://gcc.gnu.org/onlinedocs/libstdc++/latest-doxygen \
-           ./docs/doxygen-tags/opencv.tag=http://docs.opencv.org/3.4.2
+           ./docs/doxygen-tags/opencv.tag=http://docs.opencv.org/3.4.3
 @endcode
 
 Doxygen can now use the information from the tag file to link to the OpenCV

modules/core/misc/java/src/java/core+MatOfRotatedRect.java

@@ -0,0 +1,86 @@
+package org.opencv.core;
+
+import java.util.Arrays;
+import java.util.List;
+
+import org.opencv.core.RotatedRect;
+
+
+
+public class MatOfRotatedRect extends Mat {
+    // 32FC5
+    private static final int _depth = CvType.CV_32F;
+    private static final int _channels = 5;
+
+    public MatOfRotatedRect() {
+        super();
+    }
+
+    protected MatOfRotatedRect(long addr) {
+        super(addr);
+        if( !empty() && checkVector(_channels, _depth) < 0 )
+            throw new IllegalArgumentException("Incompatible Mat");
+        //FIXME: do we need release() here?
+    }
+
+    public static MatOfRotatedRect fromNativeAddr(long addr) {
+        return new MatOfRotatedRect(addr);
+    }
+
+    public MatOfRotatedRect(Mat m) {
+        super(m, Range.all());
+        if( !empty() && checkVector(_channels, _depth) < 0 )
+            throw new IllegalArgumentException("Incompatible Mat");
+        //FIXME: do we need release() here?
+    }
+
+    public MatOfRotatedRect(RotatedRect...a) {
+        super();
+        fromArray(a);
+    }
+
+    public void alloc(int elemNumber) {
+        if(elemNumber>0)
+            super.create(elemNumber, 1, CvType.makeType(_depth, _channels));
+    }
+
+    public void fromArray(RotatedRect...a) {
+        if(a==null || a.length==0)
+            return;
+        int num = a.length;
+        alloc(num);
+        float buff[] = new float[num * _channels];
+        for(int i=0; i<num; i++) {
+            RotatedRect r = a[i];
+            buff[_channels*i+0] = (float) r.center.x;
+            buff[_channels*i+1] = (float) r.center.y;
+            buff[_channels*i+2] = (float) r.size.width;
+            buff[_channels*i+3] = (float) r.size.height;
+            buff[_channels*i+4] = (float) r.angle;
+        }
+        put(0, 0, buff); //TODO: check ret val!
+    }
+
+    public RotatedRect[] toArray() {
+        int num = (int) total();
+        RotatedRect[] a = new RotatedRect[num];
+        if(num == 0)
+            return a;
+        float buff[] = new float[_channels];
+        for(int i=0; i<num; i++) {
+            get(i, 0, buff); //TODO: check ret val!
+            a[i] = new RotatedRect(new Point(buff[0],buff[1]),new Size(buff[2],buff[3]),buff[4]);
+        }
+        return a;
+    }
+
+    public void fromList(List<RotatedRect> lr) {
+        RotatedRect ap[] = lr.toArray(new RotatedRect[0]);
+        fromArray(ap);
+    }
+
+    public List<RotatedRect> toList() {
+        RotatedRect[] ar = toArray();
+        return Arrays.asList(ar);
+    }
+}

modules/core/misc/java/test/RotatedRectTest.java

@@ -1,11 +1,16 @@
 package org.opencv.test.core;
 
+import org.opencv.core.CvType;
 import org.opencv.core.Point;
 import org.opencv.core.Rect;
 import org.opencv.core.RotatedRect;
+import org.opencv.core.MatOfRotatedRect;
 import org.opencv.core.Size;
 import org.opencv.test.OpenCVTestCase;
 
+import java.util.Arrays;
+import java.util.List;
+
 public class RotatedRectTest extends OpenCVTestCase {
 
     private double angle;
@@ -188,4 +193,21 @@ public class RotatedRectTest extends OpenCVTestCase {
         assertEquals(expected, actual);
     }
 
+    public void testMatOfRotatedRect() {
+        RotatedRect a = new RotatedRect(new Point(1,2),new Size(3,4),5.678);
+        RotatedRect b = new RotatedRect(new Point(9,8),new Size(7,6),5.432);
+        MatOfRotatedRect m = new MatOfRotatedRect(a,b,a,b,a,b,a,b);
+        assertEquals(m.rows(), 8);
+        assertEquals(m.cols(), 1);
+        assertEquals(m.type(), CvType.CV_32FC(5));
+        RotatedRect[] arr = m.toArray();
+        assertEquals(arr[2].angle, a.angle, EPS);
+        assertEquals(arr[3].center.x, b.center.x);
+        assertEquals(arr[3].size.width, b.size.width);
+        List<RotatedRect> li = m.toList();
+        assertEquals(li.size(), 8);
+        RotatedRect rr = li.get(7);
+        assertEquals(rr.angle, b.angle, EPS);
+        assertEquals(rr.center.y, b.center.y);
+    }
 }

modules/core/src/va_intel.cpp

@@ -324,6 +324,163 @@ static void copy_convert_bgr_to_nv12(const VAImage& image, const Mat& bgr, unsig
         dstUV += dstStepUV;
     }
 }
+
+
+static void copy_convert_yv12_to_bgr(const VAImage& image, const unsigned char* buffer, Mat& bgr)
+{
+    const float d1 = 16.0f;
+    const float d2 = 128.0f;
+
+    static const float coeffs[5] =
+        {
+            1.163999557f,
+            2.017999649f,
+            -0.390999794f,
+            -0.812999725f,
+            1.5959997177f
+        };
+
+    CV_CheckEQ(image.format.fourcc, VA_FOURCC_YV12, "Unexpected image format");
+    CV_CheckEQ(image.num_planes, 3, "");
+
+    const size_t srcOffsetY = image.offsets[0];
+    const size_t srcOffsetV = image.offsets[1];
+    const size_t srcOffsetU = image.offsets[2];
+
+    const size_t srcStepY = image.pitches[0];
+    const size_t srcStepU = image.pitches[1];
+    const size_t srcStepV = image.pitches[2];
+
+    const size_t dstStep = bgr.step;
+
+    const unsigned char* srcY_ = buffer + srcOffsetY;
+    const unsigned char* srcV_ = buffer + srcOffsetV;
+    const unsigned char* srcU_ = buffer + srcOffsetU;
+
+    for (int y = 0; y < bgr.rows; y += 2)
+    {
+        const unsigned char* srcY0 = srcY_ + (srcStepY) * y;
+        const unsigned char* srcY1 = srcY0 + srcStepY;
+
+        const unsigned char* srcV = srcV_ + (srcStepV) * y / 2;
+        const unsigned char* srcU = srcU_ + (srcStepU) * y / 2;
+
+        unsigned char* dst0 = bgr.data + (dstStep) * y;
+        unsigned char* dst1 = dst0 + dstStep;
+
+        for (int x = 0; x < bgr.cols; x += 2)
+        {
+            float Y0 = float(srcY0[x+0]);
+            float Y1 = float(srcY0[x+1]);
+            float Y2 = float(srcY1[x+0]);
+            float Y3 = float(srcY1[x+1]);
+
+            float U = float(srcU[x/2]) - d2;
+            float V = float(srcV[x/2]) - d2;
+
+            Y0 = std::max(0.0f, Y0 - d1) * coeffs[0];
+            Y1 = std::max(0.0f, Y1 - d1) * coeffs[0];
+            Y2 = std::max(0.0f, Y2 - d1) * coeffs[0];
+            Y3 = std::max(0.0f, Y3 - d1) * coeffs[0];
+
+            float ruv = coeffs[4]*V;
+            float guv = coeffs[3]*V + coeffs[2]*U;
+            float buv = coeffs[1]*U;
+
+            dst0[(x+0)*NCHANNELS+0] = saturate_cast<unsigned char>(Y0 + buv);
+            dst0[(x+0)*NCHANNELS+1] = saturate_cast<unsigned char>(Y0 + guv);
+            dst0[(x+0)*NCHANNELS+2] = saturate_cast<unsigned char>(Y0 + ruv);
+
+            dst0[(x+1)*NCHANNELS+0] = saturate_cast<unsigned char>(Y1 + buv);
+            dst0[(x+1)*NCHANNELS+1] = saturate_cast<unsigned char>(Y1 + guv);
+            dst0[(x+1)*NCHANNELS+2] = saturate_cast<unsigned char>(Y1 + ruv);
+
+            dst1[(x+0)*NCHANNELS+0] = saturate_cast<unsigned char>(Y2 + buv);
+            dst1[(x+0)*NCHANNELS+1] = saturate_cast<unsigned char>(Y2 + guv);
+            dst1[(x+0)*NCHANNELS+2] = saturate_cast<unsigned char>(Y2 + ruv);
+
+            dst1[(x+1)*NCHANNELS+0] = saturate_cast<unsigned char>(Y3 + buv);
+            dst1[(x+1)*NCHANNELS+1] = saturate_cast<unsigned char>(Y3 + guv);
+            dst1[(x+1)*NCHANNELS+2] = saturate_cast<unsigned char>(Y3 + ruv);
+        }
+    }
+}
+
+static void copy_convert_bgr_to_yv12(const VAImage& image, const Mat& bgr, unsigned char* buffer)
+{
+    const float d1 = 16.0f;
+    const float d2 = 128.0f;
+
+    static const float coeffs[8] =
+        {
+            0.256999969f,  0.50399971f,   0.09799957f,   -0.1479988098f,
+            -0.2909994125f, 0.438999176f, -0.3679990768f, -0.0709991455f
+        };
+
+    CV_CheckEQ(image.format.fourcc, VA_FOURCC_YV12, "Unexpected image format");
+    CV_CheckEQ(image.num_planes, 3, "");
+
+    const size_t dstOffsetY = image.offsets[0];
+    const size_t dstOffsetV = image.offsets[1];
+    const size_t dstOffsetU = image.offsets[2];
+
+    const size_t dstStepY = image.pitches[0];
+    const size_t dstStepU = image.pitches[1];
+    const size_t dstStepV = image.pitches[2];
+
+    unsigned char* dstY_ = buffer + dstOffsetY;
+    unsigned char* dstV_ = buffer + dstOffsetV;
+    unsigned char* dstU_ = buffer + dstOffsetU;
+
+    const size_t srcStep = bgr.step;
+
+    for (int y = 0; y < bgr.rows; y += 2)
+    {
+        unsigned char* dstY0 = dstY_ + (dstStepY) * y;
+        unsigned char* dstY1 = dstY0 + dstStepY;
+
+        unsigned char* dstV = dstV_ + (dstStepV) * y / 2;
+        unsigned char* dstU = dstU_ + (dstStepU) * y / 2;
+
+        const unsigned char* src0 = bgr.data + (srcStep) * y;
+        const unsigned char* src1 = src0 + srcStep;
+
+        for (int x = 0; x < bgr.cols; x += 2)
+        {
+            float B0 = float(src0[(x+0)*NCHANNELS+0]);
+            float G0 = float(src0[(x+0)*NCHANNELS+1]);
+            float R0 = float(src0[(x+0)*NCHANNELS+2]);
+
+            float B1 = float(src0[(x+1)*NCHANNELS+0]);
+            float G1 = float(src0[(x+1)*NCHANNELS+1]);
+            float R1 = float(src0[(x+1)*NCHANNELS+2]);
+
+            float B2 = float(src1[(x+0)*NCHANNELS+0]);
+            float G2 = float(src1[(x+0)*NCHANNELS+1]);
+            float R2 = float(src1[(x+0)*NCHANNELS+2]);
+
+            float B3 = float(src1[(x+1)*NCHANNELS+0]);
+            float G3 = float(src1[(x+1)*NCHANNELS+1]);
+            float R3 = float(src1[(x+1)*NCHANNELS+2]);
+
+            float Y0 = coeffs[0]*R0 + coeffs[1]*G0 + coeffs[2]*B0 + d1;
+            float Y1 = coeffs[0]*R1 + coeffs[1]*G1 + coeffs[2]*B1 + d1;
+            float Y2 = coeffs[0]*R2 + coeffs[1]*G2 + coeffs[2]*B2 + d1;
+            float Y3 = coeffs[0]*R3 + coeffs[1]*G3 + coeffs[2]*B3 + d1;
+
+            float U = coeffs[3]*R0 + coeffs[4]*G0 + coeffs[5]*B0 + d2;
+            float V = coeffs[5]*R0 + coeffs[6]*G0 + coeffs[7]*B0 + d2;
+
+            dstY0[x+0] = saturate_cast<unsigned char>(Y0);
+            dstY0[x+1] = saturate_cast<unsigned char>(Y1);
+            dstY1[x+0] = saturate_cast<unsigned char>(Y2);
+            dstY1[x+1] = saturate_cast<unsigned char>(Y3);
+
+            dstU[x/2] = saturate_cast<unsigned char>(U);
+            dstV[x/2] = saturate_cast<unsigned char>(V);
+        }
+    }
+}
 #endif // HAVE_VA
 
 void convertToVASurface(VADisplay display, InputArray src, VASurfaceID surface, Size size)
@@ -412,9 +569,12 @@ void convertToVASurface(VADisplay display, InputArray src, VASurfaceID surface,
         if (status != VA_STATUS_SUCCESS)
             CV_Error(cv::Error::StsError, "VA-API: vaMapBuffer failed");
 
-        CV_Assert(image.format.fourcc == VA_FOURCC_NV12);
+        if (image.format.fourcc == VA_FOURCC_NV12)
-
+            copy_convert_bgr_to_nv12(image, m, buffer);
-        copy_convert_bgr_to_nv12(image, m, buffer);
+        if (image.format.fourcc == VA_FOURCC_YV12)
+            copy_convert_bgr_to_yv12(image, m, buffer);
+        else
+            CV_Check((int)image.format.fourcc, image.format.fourcc == VA_FOURCC_NV12 || image.format.fourcc == VA_FOURCC_YV12, "Unexpected image format");
 
         status = vaUnmapBuffer(display, image.buf);
         if (status != VA_STATUS_SUCCESS)
@@ -510,9 +670,12 @@ void convertFromVASurface(VADisplay display, VASurfaceID surface, Size size, Out
         if (status != VA_STATUS_SUCCESS)
             CV_Error(cv::Error::StsError, "VA-API: vaMapBuffer failed");
 
-        CV_Assert(image.format.fourcc == VA_FOURCC_NV12);
+        if (image.format.fourcc == VA_FOURCC_NV12)
-
+            copy_convert_nv12_to_bgr(image, buffer, m);
-        copy_convert_nv12_to_bgr(image, buffer, m);
+        if (image.format.fourcc == VA_FOURCC_YV12)
+            copy_convert_yv12_to_bgr(image, buffer, m);
+        else
+            CV_Check((int)image.format.fourcc, image.format.fourcc == VA_FOURCC_NV12 || image.format.fourcc == VA_FOURCC_YV12, "Unexpected image format");
 
         status = vaUnmapBuffer(display, image.buf);
         if (status != VA_STATUS_SUCCESS)

modules/core/test/test_arithm.cpp

@@ -2158,4 +2158,71 @@ TEST(Core_Norm, IPP_regression_NORM_L1_16UC3_small)
     EXPECT_EQ((double)20*cn, cv::norm(a, b, NORM_L1, mask));
 }
 
+
+TEST(Core_ConvertTo, regression_12121)
+{
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(-1));
+        Mat dst;
+        src.convertTo(dst, CV_8U);
+        EXPECT_EQ(0, dst.at<uchar>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN));
+        Mat dst;
+        src.convertTo(dst, CV_8U);
+        EXPECT_EQ(0, dst.at<uchar>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN + 32767));
+        Mat dst;
+        src.convertTo(dst, CV_8U);
+        EXPECT_EQ(0, dst.at<uchar>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN + 32768));
+        Mat dst;
+        src.convertTo(dst, CV_8U);
+        EXPECT_EQ(0, dst.at<uchar>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(32768));
+        Mat dst;
+        src.convertTo(dst, CV_8U);
+        EXPECT_EQ(255, dst.at<uchar>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN));
+        Mat dst;
+        src.convertTo(dst, CV_16U);
+        EXPECT_EQ(0, dst.at<ushort>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN + 32767));
+        Mat dst;
+        src.convertTo(dst, CV_16U);
+        EXPECT_EQ(0, dst.at<ushort>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(INT_MIN + 32768));
+        Mat dst;
+        src.convertTo(dst, CV_16U);
+        EXPECT_EQ(0, dst.at<ushort>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+
+    {
+        Mat src(4, 64, CV_32SC1, Scalar(65536));
+        Mat dst;
+        src.convertTo(dst, CV_16U);
+        EXPECT_EQ(65535, dst.at<ushort>(0, 0)) << "src=" << src.at<int>(0, 0);
+    }
+}
+
 }} // namespace

modules/cudaarithm/src/cuda/math.cu

@@ -278,20 +278,12 @@ namespace
 {
     template<typename T, bool Signed = numeric_limits<T>::is_signed> struct PowOp : unary_function<T, T>
     {
-        float power;
+        typedef typename LargerType<T, float>::type LargerType;
+        LargerType power;
 
         __device__ __forceinline__ T operator()(T e) const
         {
-            return cudev::saturate_cast<T>(__powf((float)e, power));
+            T res = cudev::saturate_cast<T>(__powf(e < 0 ? -e : e, power));
-        }
-    };
-    template<typename T> struct PowOp<T, true> : unary_function<T, T>
-    {
-        float power;
-
-        __device__ __forceinline__ T operator()(T e) const
-        {
-            T res = cudev::saturate_cast<T>(__powf((float)e, power));
 
             if ((e < 0) && (1 & static_cast<int>(power)))
                 res *= -1;
@@ -299,22 +291,15 @@ namespace
             return res;
         }
     };
-    template<> struct PowOp<float> : unary_function<float, float>
-    {
-        float power;
 
-        __device__ __forceinline__ float operator()(float e) const
+    template<typename T> struct PowOp<T, false> : unary_function<T, T>
-        {
-            return __powf(::fabs(e), power);
-        }
-    };
-    template<> struct PowOp<double> : unary_function<double, double>
     {
-        double power;
+        typedef typename LargerType<T, float>::type LargerType;
+        LargerType power;
 
-        __device__ __forceinline__ double operator()(double e) const
+        __device__ __forceinline__ T operator()(T e) const
         {
-            return ::pow(::fabs(e), power);
+            return cudev::saturate_cast<T>(__powf(e, power));
         }
     };
 

modules/dnn/include/opencv2/dnn/dnn.hpp

@@ -46,9 +46,9 @@
 #include <opencv2/core.hpp>
 
 #if !defined CV_DOXYGEN && !defined CV_DNN_DONT_ADD_EXPERIMENTAL_NS
-#define CV__DNN_EXPERIMENTAL_NS_BEGIN namespace experimental_dnn_v6 {
+#define CV__DNN_EXPERIMENTAL_NS_BEGIN namespace experimental_dnn_34_v7 {
 #define CV__DNN_EXPERIMENTAL_NS_END }
-namespace cv { namespace dnn { namespace experimental_dnn_v6 { } using namespace experimental_dnn_v6; }}
+namespace cv { namespace dnn { namespace experimental_dnn_34_v7 { } using namespace experimental_dnn_34_v7; }}
 #else
 #define CV__DNN_EXPERIMENTAL_NS_BEGIN
 #define CV__DNN_EXPERIMENTAL_NS_END
@@ -900,7 +900,7 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
                                CV_OUT std::vector<int>& indices,
                                const float eta = 1.f, const int top_k = 0);
 
-    CV_EXPORTS void NMSBoxes(const std::vector<RotatedRect>& bboxes, const std::vector<float>& scores,
+    CV_EXPORTS_AS(NMSBoxesRotated) void NMSBoxes(const std::vector<RotatedRect>& bboxes, const std::vector<float>& scores,
                              const float score_threshold, const float nms_threshold,
                              CV_OUT std::vector<int>& indices,
                              const float eta = 1.f, const int top_k = 0);

modules/dnn/src/dnn.cpp

@@ -699,9 +699,9 @@ public:
         }
     }
 
-    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool forceCreate, bool use_half)
+    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool use_half)
     {
-        if (!DNN_DISABLE_MEMORY_OPTIMIZATIONS && !forceCreate)
+        if (!DNN_DISABLE_MEMORY_OPTIMIZATIONS)
         {
             Mat bestBlob;
             LayerPin bestBlobPin;
@@ -747,7 +747,7 @@ public:
 
     void allocateBlobsForLayer(LayerData &ld, const LayerShapes& layerShapes,
                                std::vector<LayerPin>& pinsForInternalBlobs,
-                               bool forceCreate = false, bool use_half = false)
+                               bool use_half = false)
     {
         CV_TRACE_FUNCTION();
 
@@ -818,7 +818,7 @@ public:
                         reuse(ld.inputBlobsId[0], blobPin);
                     }
                     else
-                        reuseOrCreate(shapes[index], blobPin, *blobs[index], forceCreate, use_half);
+                        reuseOrCreate(shapes[index], blobPin, *blobs[index], use_half);
                 }
             }
         }
@@ -1607,7 +1607,6 @@ struct Net::Impl
 
         std::vector<LayerPin> pinsForInternalBlobs;
         blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs,
-                                          preferableBackend == DNN_BACKEND_INFERENCE_ENGINE,
                                           preferableBackend == DNN_BACKEND_OPENCV &&
                                           preferableTarget == DNN_TARGET_OPENCL_FP16);
         ld.outputBlobsWrappers.resize(ld.outputBlobs.size());

modules/dnn/src/layers/convolution_layer.cpp

@@ -81,6 +81,7 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
+#ifdef HAVE_INF_ENGINE
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
         {
             if (type == "Convolution")
@@ -91,13 +92,19 @@ public:
                 const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
                 const int group = numOutput / outGroupCn;
                 if (group != 1)
+                {
+#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R3)
+                    return preferableTarget == DNN_TARGET_CPU;
+#endif
                     return false;
+                }
                 if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
                     return dilation.width == 1 && dilation.height == 1;
                 return true;
             }
         }
         else
+#endif  // HAVE_INF_ENGINE
             return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
     }
 

modules/dnn/src/layers/crop_and_resize_layer.cpp

@@ -99,6 +99,13 @@ public:
                 }
             }
         }
+        if (boxes.rows < out.size[0])
+        {
+            // left = top = right = bottom = 0
+            std::vector<cv::Range> dstRanges(4, Range::all());
+            dstRanges[0] = Range(boxes.rows, out.size[0]);
+            out(dstRanges).setTo(inp.ptr<float>(0, 0, 0)[0]);
+        }
     }
 
 private:

modules/dnn/src/layers/detection_output_layer.cpp

@@ -115,6 +115,7 @@ public:
     // It's true whenever predicted bounding boxes and proposals are normalized to [0, 1].
     bool _bboxesNormalized;
     bool _clip;
+    bool _groupByClasses;
 
     enum { _numAxes = 4 };
     static const std::string _layerName;
@@ -183,6 +184,7 @@ public:
         _locPredTransposed = getParameter<bool>(params, "loc_pred_transposed", 0, false, false);
         _bboxesNormalized = getParameter<bool>(params, "normalized_bbox", 0, false, true);
         _clip = getParameter<bool>(params, "clip", 0, false, false);
+        _groupByClasses = getParameter<bool>(params, "group_by_classes", 0, false, true);
 
         getCodeType(params);
 
@@ -381,7 +383,7 @@ public:
             {
                 count += outputDetections_(i, &outputsData[count * 7],
                                            allDecodedBBoxes[i], allConfidenceScores[i],
-                                           allIndices[i]);
+                                           allIndices[i], _groupByClasses);
             }
             CV_Assert(count == numKept);
         }
@@ -497,7 +499,7 @@ public:
         {
             count += outputDetections_(i, &outputsData[count * 7],
                                        allDecodedBBoxes[i], allConfidenceScores[i],
-                                       allIndices[i]);
+                                       allIndices[i], _groupByClasses);
         }
         CV_Assert(count == numKept);
     }
@@ -505,9 +507,36 @@ public:
     size_t outputDetections_(
             const int i, float* outputsData,
             const LabelBBox& decodeBBoxes, Mat& confidenceScores,
-            const std::map<int, std::vector<int> >& indicesMap
+            const std::map<int, std::vector<int> >& indicesMap,
+            bool groupByClasses
     )
     {
+        std::vector<int> dstIndices;
+        std::vector<std::pair<float, int> > allScores;
+        for (std::map<int, std::vector<int> >::const_iterator it = indicesMap.begin(); it != indicesMap.end(); ++it)
+        {
+            int label = it->first;
+            if (confidenceScores.rows <= label)
+                CV_Error_(cv::Error::StsError, ("Could not find confidence predictions for label %d", label));
+            const std::vector<float>& scores = confidenceScores.row(label);
+            const std::vector<int>& indices = it->second;
+
+            const int numAllScores = allScores.size();
+            allScores.reserve(numAllScores + indices.size());
+            for (size_t j = 0; j < indices.size(); ++j)
+            {
+                allScores.push_back(std::make_pair(scores[indices[j]], numAllScores + j));
+            }
+        }
+        if (!groupByClasses)
+            std::sort(allScores.begin(), allScores.end(), util::SortScorePairDescend<int>);
+
+        dstIndices.resize(allScores.size());
+        for (size_t j = 0; j < dstIndices.size(); ++j)
+        {
+            dstIndices[allScores[j].second] = j;
+        }
+
         size_t count = 0;
         for (std::map<int, std::vector<int> >::const_iterator it = indicesMap.begin(); it != indicesMap.end(); ++it)
         {
@@ -524,14 +553,15 @@ public:
             for (size_t j = 0; j < indices.size(); ++j, ++count)
             {
                 int idx = indices[j];
+                int dstIdx = dstIndices[count];
                 const util::NormalizedBBox& decode_bbox = label_bboxes->second[idx];
-                outputsData[count * 7] = i;
+                outputsData[dstIdx * 7] = i;
-                outputsData[count * 7 + 1] = label;
+                outputsData[dstIdx * 7 + 1] = label;
-                outputsData[count * 7 + 2] = scores[idx];
+                outputsData[dstIdx * 7 + 2] = scores[idx];
-                outputsData[count * 7 + 3] = decode_bbox.xmin;
+                outputsData[dstIdx * 7 + 3] = decode_bbox.xmin;
-                outputsData[count * 7 + 4] = decode_bbox.ymin;
+                outputsData[dstIdx * 7 + 4] = decode_bbox.ymin;
-                outputsData[count * 7 + 5] = decode_bbox.xmax;
+                outputsData[dstIdx * 7 + 5] = decode_bbox.xmax;
-                outputsData[count * 7 + 6] = decode_bbox.ymax;
+                outputsData[dstIdx * 7 + 6] = decode_bbox.ymax;
             }
         }
         return count;

modules/dnn/src/layers/elementwise_layers.cpp

@@ -599,7 +599,8 @@ struct ELUFunctor
 
     bool supportBackend(int backendId, int)
     {
-        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE;
     }
 
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
@@ -653,8 +654,8 @@ struct ELUFunctor
 #ifdef HAVE_INF_ENGINE
     InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
     {
-        CV_Error(Error::StsNotImplemented, "ELU");
+        lp.type = "ELU";
-        return InferenceEngine::CNNLayerPtr();
+        return InferenceEngine::CNNLayerPtr(new InferenceEngine::CNNLayer(lp));
     }
 #endif  // HAVE_INF_ENGINE
 

modules/dnn/src/layers/lrn_layer.cpp

@@ -91,8 +91,8 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
         return backendId == DNN_BACKEND_OPENCV ||
-               backendId == DNN_BACKEND_HALIDE && haveHalide() ||
+               backendId == DNN_BACKEND_HALIDE ||
-               backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine();
+               backendId == DNN_BACKEND_INFERENCE_ENGINE && (preferableTarget != DNN_TARGET_MYRIAD || type == CHANNEL_NRM);
     }
 
 #ifdef HAVE_OPENCL

modules/dnn/src/layers/resize_layer.cpp

@@ -33,9 +33,7 @@ public:
         interpolation = params.get<String>("interpolation");
         CV_Assert(interpolation == "nearest" || interpolation == "bilinear");
 
-        bool alignCorners = params.get<bool>("align_corners", false);
+        alignCorners = params.get<bool>("align_corners", false);
-        if (alignCorners)
-            CV_Error(Error::StsNotImplemented, "Resize with align_corners=true is not implemented");
     }
 
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
@@ -66,8 +64,15 @@ public:
             outHeight = outputs[0].size[2];
             outWidth = outputs[0].size[3];
         }
-        scaleHeight = static_cast<float>(inputs[0]->size[2]) / outHeight;
+        if (alignCorners && outHeight > 1)
-        scaleWidth = static_cast<float>(inputs[0]->size[3]) / outWidth;
+            scaleHeight = static_cast<float>(inputs[0]->size[2] - 1) / (outHeight - 1);
+        else
+            scaleHeight = static_cast<float>(inputs[0]->size[2]) / outHeight;
+
+        if (alignCorners && outWidth > 1)
+            scaleWidth = static_cast<float>(inputs[0]->size[3] - 1) / (outWidth - 1);
+        else
+            scaleWidth = static_cast<float>(inputs[0]->size[3]) / outWidth;
     }
 
     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
@@ -166,6 +171,7 @@ protected:
     int outWidth, outHeight, zoomFactorWidth, zoomFactorHeight;
     String interpolation;
     float scaleWidth, scaleHeight;
+    bool alignCorners;
 };
 
 

modules/dnn/src/op_inf_engine.hpp

@@ -24,6 +24,7 @@
 
 #define INF_ENGINE_RELEASE_2018R1 2018010000
 #define INF_ENGINE_RELEASE_2018R2 2018020000
+#define INF_ENGINE_RELEASE_2018R3 2018030000
 
 #ifndef INF_ENGINE_RELEASE
 #warning("IE version have not been provided via command-line. Using 2018R2 by default")
@@ -31,6 +32,7 @@
 #endif
 
 #define INF_ENGINE_VER_MAJOR_GT(ver) (((INF_ENGINE_RELEASE) / 10000) > ((ver) / 10000))
+#define INF_ENGINE_VER_MAJOR_GE(ver) (((INF_ENGINE_RELEASE) / 10000) >= ((ver) / 10000))
 
 #endif  // HAVE_INF_ENGINE
 

modules/dnn/src/tensorflow/tf_importer.cpp

@@ -737,11 +737,18 @@ void TFImporter::populateNet(Net dstNet)
         int predictedLayout = predictOutputDataLayout(net, layer, data_layouts);
         data_layouts[name] = predictedLayout;
 
-        if (type == "Conv2D" || type == "SpaceToBatchND" || type == "DepthwiseConv2dNative")
+        if (type == "Conv2D" || type == "SpaceToBatchND" || type == "DepthwiseConv2dNative" || type == "Pad")
         {
             // The first node of dilated convolution subgraph.
             // Extract input node, dilation rate and paddings.
             std::string input = layer.input(0);
+            StrIntVector next_layers;
+            if (type == "SpaceToBatchND" || type == "Pad")
+            {
+                next_layers = getNextLayers(net, name, "Conv2D");
+                if (next_layers.empty())
+                    next_layers = getNextLayers(net, name, "DepthwiseConv2dNative");
+            }
             if (type == "SpaceToBatchND")
             {
                 // op: "SpaceToBatchND"
@@ -762,17 +769,57 @@ void TFImporter::populateNet(Net dstNet)
                 layerParams.set("pad_h", paddings.at<float>(0));
                 layerParams.set("pad_w", paddings.at<float>(2));
 
-                StrIntVector next_layers = getNextLayers(net, name, "Conv2D");
-                if (next_layers.empty())
-                {
-                    next_layers = getNextLayers(net, name, "DepthwiseConv2dNative");
-                }
                 CV_Assert(next_layers.size() == 1);
                 layer = net.node(next_layers[0].second);
                 layers_to_ignore.insert(next_layers[0].first);
                 name = layer.name();
                 type = layer.op();
             }
+            else if (type == "Pad")
+            {
+                Mat paddings = getTensorContent(getConstBlob(layer, value_id, 1));
+                CV_Assert(paddings.type() == CV_32SC1);
+                if (paddings.total() == 8)
+                {
+                    // Perhabs, we have NHWC padding dimensions order.
+                    //  N    H    W    C
+                    // 0 1  2 3  4 5  6 7
+                    std::swap(paddings.at<int32_t>(2), paddings.at<int32_t>(6));
+                    std::swap(paddings.at<int32_t>(3), paddings.at<int32_t>(7));
+                    //  N    C    W    H
+                    // 0 1  2 3  4 5  6 7
+                    std::swap(paddings.at<int32_t>(4), paddings.at<int32_t>(6));
+                    std::swap(paddings.at<int32_t>(5), paddings.at<int32_t>(7));
+                    //  N    C    H    W
+                    // 0 1  2 3  4 5  6 7
+                }
+                if (next_layers.empty() || paddings.total() != 8 ||
+                    paddings.at<int32_t>(4) != paddings.at<int32_t>(5) ||
+                    paddings.at<int32_t>(6) != paddings.at<int32_t>(7))
+                {
+                    // Just a single padding layer.
+                    layerParams.set("paddings", DictValue::arrayInt<int*>((int*)paddings.data, paddings.total()));
+
+                    int id = dstNet.addLayer(name, "Padding", layerParams);
+                    layer_id[name] = id;
+
+                    connect(layer_id, dstNet, parsePin(input), id, 0);
+                    continue;
+                }
+                else
+                {
+                    // Merge with subsequent convolutional layer.
+                    CV_Assert(next_layers.size() == 1);
+
+                    layerParams.set("pad_h", paddings.at<int32_t>(4));
+                    layerParams.set("pad_w", paddings.at<int32_t>(6));
+
+                    layer = net.node(next_layers[0].second);
+                    layers_to_ignore.insert(next_layers[0].first);
+                    name = layer.name();
+                    type = layer.op();
+                }
+            }
 
             // For the object detection networks, TensorFlow Object Detection API
             // predicts deltas for bounding boxes in yxYX (ymin, xmin, ymax, xmax)
@@ -784,7 +831,7 @@ void TFImporter::populateNet(Net dstNet)
             layerParams.set("bias_term", false);
             layerParams.blobs.resize(1);
 
-            StrIntVector next_layers = getNextLayers(net, name, "BiasAdd");
+            next_layers = getNextLayers(net, name, "BiasAdd");
             if (next_layers.size() == 1) {
                 layerParams.set("bias_term", true);
                 layerParams.blobs.resize(2);
@@ -1416,31 +1463,6 @@ void TFImporter::populateNet(Net dstNet)
                 }
             }
         }
-        else if (type == "Pad")
-        {
-            Mat paddings = getTensorContent(getConstBlob(layer, value_id, 1));
-            CV_Assert(paddings.type() == CV_32SC1);
-            if (paddings.total() == 8)
-            {
-                // Perhabs, we have NHWC padding dimensions order.
-                //  N    H    W    C
-                // 0 1  2 3  4 5  6 7
-                std::swap(*paddings.ptr<int32_t>(0, 2), *paddings.ptr<int32_t>(0, 6));
-                std::swap(*paddings.ptr<int32_t>(0, 3), *paddings.ptr<int32_t>(0, 7));
-                //  N    C    W    H
-                // 0 1  2 3  4 5  6 7
-                std::swap(*paddings.ptr<int32_t>(0, 4), *paddings.ptr<int32_t>(0, 6));
-                std::swap(*paddings.ptr<int32_t>(0, 5), *paddings.ptr<int32_t>(0, 7));
-                //  N    C    H    W
-                // 0 1  2 3  4 5  6 7
-            }
-            layerParams.set("paddings", DictValue::arrayInt<int*>((int*)paddings.data, paddings.total()));
-
-            int id = dstNet.addLayer(name, "Padding", layerParams);
-            layer_id[name] = id;
-
-            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
-        }
         else if (type == "FusedBatchNorm")
         {
             // op: "FusedBatchNorm"

modules/dnn/test/test_backends.cpp

@@ -222,9 +222,12 @@ TEST_P(DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages)
 
 TEST_P(DNNTestNetwork, OpenFace)
 {
+#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE < 2018030000
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("Test is enabled starts from OpenVINO 2018R3");
+#endif
     if (backend == DNN_BACKEND_HALIDE ||
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16))
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
         throw SkipTestException("");
     processNet("dnn/openface_nn4.small2.v1.t7", "", Size(96, 96), "");
 }
@@ -253,12 +256,19 @@ TEST_P(DNNTestNetwork, Inception_v2_SSD_TensorFlow)
 
 TEST_P(DNNTestNetwork, DenseNet_121)
 {
-    if ((backend == DNN_BACKEND_HALIDE) ||
+    if (backend == DNN_BACKEND_HALIDE)
-        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ||
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && (target == DNN_TARGET_OPENCL_FP16 ||
-                                                     target == DNN_TARGET_MYRIAD)))
         throw SkipTestException("");
-    processNet("dnn/DenseNet_121.caffemodel", "dnn/DenseNet_121.prototxt", Size(224, 224), "", "caffe");
+
+    float l1 = 0.0, lInf = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        l1 = 9e-3; lInf = 5e-2;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 6e-2; lInf = 0.27;
+    }
+    processNet("dnn/DenseNet_121.caffemodel", "dnn/DenseNet_121.prototxt", Size(224, 224), "", "", l1, lInf);
 }
 
 TEST_P(DNNTestNetwork, FastNeuralStyle_eccv16)

modules/dnn/test/test_caffe_importer.cpp

@@ -374,14 +374,6 @@ TEST(Reproducibility_GoogLeNet_fp16, Accuracy)
 TEST_P(Test_Caffe_nets, Colorization)
 {
     checkBackend();
-    if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD) ||
-        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
-        throw SkipTestException("");
-
-    const float l1 = 4e-4;
-    const float lInf = 3e-3;
-
     Mat inp = blobFromNPY(_tf("colorization_inp.npy"));
     Mat ref = blobFromNPY(_tf("colorization_out.npy"));
     Mat kernel = blobFromNPY(_tf("colorization_pts_in_hull.npy"));
@@ -398,11 +390,15 @@ TEST_P(Test_Caffe_nets, Colorization)
     net.setInput(inp);
     Mat out = net.forward();
 
+    // Reference output values are in range [-29.1, 69.5]
+    const double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.21 : 4e-4;
+    const double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 5.3 : 3e-3;
     normAssert(out, ref, "", l1, lInf);
 }
 
-TEST(Reproducibility_DenseNet_121, Accuracy)
+TEST_P(Test_Caffe_nets, DenseNet_121)
 {
+    checkBackend();
     const string proto = findDataFile("dnn/DenseNet_121.prototxt", false);
     const string model = findDataFile("dnn/DenseNet_121.caffemodel", false);
 
@@ -411,12 +407,23 @@ TEST(Reproducibility_DenseNet_121, Accuracy)
     Mat ref = blobFromNPY(_tf("densenet_121_output.npy"));
 
     Net net = readNetFromCaffe(proto, model);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
 
     net.setInput(inp);
     Mat out = net.forward();
 
-    normAssert(out, ref);
+    // Reference is an array of 1000 values from a range [-6.16, 7.9]
+    float l1 = default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        l1 = 0.017; lInf = 0.067;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.097; lInf = 0.52;
+    }
+    normAssert(out, ref, "", l1, lInf);
 }
 
 TEST(Test_Caffe, multiple_inputs)

modules/dnn/test/test_ie_models.cpp

@@ -177,7 +177,8 @@ TEST_P(DNNTestOpenVINO, models)
     Target target = (dnn::Target)(int)get<0>(GetParam());
     std::string modelName = get<1>(GetParam());
 
-    if (modelName == "semantic-segmentation-adas-0001" && target == DNN_TARGET_OPENCL_FP16)
+    if ((modelName == "semantic-segmentation-adas-0001" && target == DNN_TARGET_OPENCL_FP16) ||
+        (modelName == "vehicle-license-plate-detection-barrier-0106"))
         throw SkipTestException("");
 
     std::string precision = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? "FP16" : "FP32";

modules/dnn/test/test_layers.cpp

@@ -127,15 +127,9 @@ TEST_P(Test_Caffe_layers, Softmax)
     testLayerUsingCaffeModels("layer_softmax");
 }
 
-TEST_P(Test_Caffe_layers, LRN_spatial)
+TEST_P(Test_Caffe_layers, LRN)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
-        throw SkipTestException("");
     testLayerUsingCaffeModels("layer_lrn_spatial");
-}
-
-TEST_P(Test_Caffe_layers, LRN_channels)
-{
     testLayerUsingCaffeModels("layer_lrn_channels");
 }
 

modules/dnn/test/test_tf_importer.cpp

@@ -399,8 +399,10 @@ TEST_P(Test_TensorFlow_nets, opencv_face_detector_uint8)
 TEST_P(Test_TensorFlow_nets, EAST_text_detection)
 {
     checkBackend();
-    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE < 2018030000
-        throw SkipTestException("");
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("Test is enabled starts from OpenVINO 2018R3");
+#endif
 
     std::string netPath = findDataFile("dnn/frozen_east_text_detection.pb", false);
     std::string imgPath = findDataFile("cv/ximgproc/sources/08.png", false);
@@ -425,8 +427,25 @@ TEST_P(Test_TensorFlow_nets, EAST_text_detection)
     Mat scores = outs[0];
     Mat geometry = outs[1];
 
-    normAssert(scores, blobFromNPY(refScoresPath), "scores");
+    // Scores are in range [0, 1]. Geometry values are in range [-0.23, 290]
-    normAssert(geometry, blobFromNPY(refGeometryPath), "geometry", 1e-4, 3e-3);
+    double l1_scores = default_l1, lInf_scores = default_lInf;
+    double l1_geometry = default_l1, lInf_geometry = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        lInf_scores = 0.11;
+        l1_geometry = 0.28; lInf_geometry = 5.94;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        lInf_scores = 0.214;
+        l1_geometry = 0.47; lInf_geometry = 15.34;
+    }
+    else
+    {
+        l1_geometry = 1e-4, lInf_geometry = 3e-3;
+    }
+    normAssert(scores, blobFromNPY(refScoresPath), "scores", l1_scores, lInf_scores);
+    normAssert(geometry, blobFromNPY(refGeometryPath), "geometry", l1_geometry, lInf_geometry);
 }
 
 INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_nets, dnnBackendsAndTargets());
@@ -537,4 +556,56 @@ TEST(Test_TensorFlow, two_inputs)
     normAssert(out, firstInput + secondInput);
 }
 
+TEST(Test_TensorFlow, Mask_RCNN)
+{
+    std::string proto = findDataFile("dnn/mask_rcnn_inception_v2_coco_2018_01_28.pbtxt", false);
+    std::string model = findDataFile("dnn/mask_rcnn_inception_v2_coco_2018_01_28.pb", false);
+
+    Net net = readNetFromTensorflow(model, proto);
+    Mat img = imread(findDataFile("dnn/street.png", false));
+    Mat refDetections = blobFromNPY(path("mask_rcnn_inception_v2_coco_2018_01_28.detection_out.npy"));
+    Mat refMasks = blobFromNPY(path("mask_rcnn_inception_v2_coco_2018_01_28.detection_masks.npy"));
+    Mat blob = blobFromImage(img, 1.0f, Size(800, 800), Scalar(), true, false);
+
+    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+
+    net.setInput(blob);
+
+    // Mask-RCNN predicts bounding boxes and segmentation masks.
+    std::vector<String> outNames(2);
+    outNames[0] = "detection_out_final";
+    outNames[1] = "detection_masks";
+
+    std::vector<Mat> outs;
+    net.forward(outs, outNames);
+
+    Mat outDetections = outs[0];
+    Mat outMasks = outs[1];
+    normAssertDetections(refDetections, outDetections, "", /*threshold for zero confidence*/1e-5);
+
+    // Output size of masks is NxCxHxW where
+    // N - number of detected boxes
+    // C - number of classes (excluding background)
+    // HxW - segmentation shape
+    const int numDetections = outDetections.size[2];
+
+    int masksSize[] = {1, numDetections, outMasks.size[2], outMasks.size[3]};
+    Mat masks(4, &masksSize[0], CV_32F);
+
+    std::vector<cv::Range> srcRanges(4, cv::Range::all());
+    std::vector<cv::Range> dstRanges(4, cv::Range::all());
+
+    outDetections = outDetections.reshape(1, outDetections.total() / 7);
+    for (int i = 0; i < numDetections; ++i)
+    {
+        // Get a class id for this bounding box and copy mask only for that class.
+        int classId = static_cast<int>(outDetections.at<float>(i, 1));
+        srcRanges[0] = dstRanges[1] = cv::Range(i, i + 1);
+        srcRanges[1] = cv::Range(classId, classId + 1);
+        outMasks(srcRanges).copyTo(masks(dstRanges));
+    }
+    cv::Range topRefMasks[] = {Range::all(), Range(0, numDetections), Range::all(), Range::all()};
+    normAssert(masks, refMasks(&topRefMasks[0]));
+}
+
 }

modules/dnn/test/test_torch_importer.cpp

@@ -242,15 +242,23 @@ TEST_P(Test_Torch_layers, net_residual)
     runTorchNet("net_residual", "", false, true);
 }
 
-typedef testing::TestWithParam<Target> Test_Torch_nets;
+class Test_Torch_nets : public DNNTestLayer {};
 
 TEST_P(Test_Torch_nets, OpenFace_accuracy)
 {
+#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE < 2018030000
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("Test is enabled starts from OpenVINO 2018R3");
+#endif
+    checkBackend();
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
+        throw SkipTestException("");
+
     const string model = findDataFile("dnn/openface_nn4.small2.v1.t7", false);
     Net net = readNetFromTorch(model);
 
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
-    net.setPreferableTarget(GetParam());
+    net.setPreferableTarget(target);
 
     Mat sample = imread(findDataFile("cv/shared/lena.png", false));
     Mat sampleF32(sample.size(), CV_32FC3);
@@ -264,11 +272,16 @@ TEST_P(Test_Torch_nets, OpenFace_accuracy)
     Mat out = net.forward();
 
     Mat outRef = readTorchBlob(_tf("net_openface_output.dat"), true);
-    normAssert(out, outRef);
+    normAssert(out, outRef, "", default_l1, default_lInf);
 }
 
 TEST_P(Test_Torch_nets, ENet_accuracy)
 {
+    checkBackend();
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
+        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
+
     Net net;
     {
         const string model = findDataFile("dnn/Enet-model-best.net", false);
@@ -276,8 +289,8 @@ TEST_P(Test_Torch_nets, ENet_accuracy)
         ASSERT_TRUE(!net.empty());
     }
 
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
-    net.setPreferableTarget(GetParam());
+    net.setPreferableTarget(target);
 
     Mat sample = imread(_tf("street.png", false));
     Mat inputBlob = blobFromImage(sample, 1./255);
@@ -314,6 +327,7 @@ TEST_P(Test_Torch_nets, ENet_accuracy)
 //   -model models/instance_norm/feathers.t7
 TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
 {
+    checkBackend();
     std::string models[] = {"dnn/fast_neural_style_eccv16_starry_night.t7",
                             "dnn/fast_neural_style_instance_norm_feathers.t7"};
     std::string targets[] = {"dnn/lena_starry_night.png", "dnn/lena_feathers.png"};
@@ -323,8 +337,8 @@ TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
         const string model = findDataFile(models[i], false);
         Net net = readNetFromTorch(model);
 
-        net.setPreferableBackend(DNN_BACKEND_OPENCV);
+        net.setPreferableBackend(backend);
-        net.setPreferableTarget(GetParam());
+        net.setPreferableTarget(target);
 
         Mat img = imread(findDataFile("dnn/googlenet_1.png", false));
         Mat inputBlob = blobFromImage(img, 1.0, Size(), Scalar(103.939, 116.779, 123.68), false);
@@ -341,12 +355,20 @@ TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
         Mat ref = imread(findDataFile(targets[i]));
         Mat refBlob = blobFromImage(ref, 1.0, Size(), Scalar(), false);
 
-        normAssert(out, refBlob, "", 0.5, 1.1);
+        if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+        {
+            double normL1 = cvtest::norm(refBlob, out, cv::NORM_L1) / refBlob.total();
+            if (target == DNN_TARGET_MYRIAD)
+                EXPECT_LE(normL1, 4.0f);
+            else
+                EXPECT_LE(normL1, 0.6f);
+        }
+        else
+            normAssert(out, refBlob, "", 0.5, 1.1);
     }
 }
 
-INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_nets, availableDnnTargets());
+INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_nets, dnnBackendsAndTargets());
-
 
 // Test a custom layer
 // https://github.com/torch/nn/blob/master/doc/convolution.md#nn.SpatialUpSamplingNearest

modules/python/src2/cv2.cpp

@@ -199,6 +199,7 @@ typedef std::vector<Vec6f> vector_Vec6f;
 typedef std::vector<Vec4i> vector_Vec4i;
 typedef std::vector<Rect> vector_Rect;
 typedef std::vector<Rect2d> vector_Rect2d;
+typedef std::vector<RotatedRect> vector_RotatedRect;
 typedef std::vector<KeyPoint> vector_KeyPoint;
 typedef std::vector<Mat> vector_Mat;
 typedef std::vector<std::vector<Mat> > vector_vector_Mat;
@@ -1643,6 +1644,18 @@ template<> struct pyopencvVecConverter<String>
     }
 };
 
+template<> struct pyopencvVecConverter<RotatedRect>
+{
+    static bool to(PyObject* obj, std::vector<RotatedRect>& value, const ArgInfo info)
+    {
+        return pyopencv_to_generic_vec(obj, value, info);
+    }
+    static PyObject* from(const std::vector<RotatedRect>& value)
+    {
+        return pyopencv_from_generic_vec(value);
+    }
+};
+
 template<>
 bool pyopencv_to(PyObject *obj, TermCriteria& dst, const char *name)
 {

modules/shape/src/aff_trans.cpp

@@ -165,8 +165,8 @@ static Mat _localAffineEstimate(const std::vector<Point2f>& shape1, const std::v
             }
             else
             {
-                therow.at<float>(0,0)=-shape1[contPt].y;
+                therow.at<float>(0,0)=shape1[contPt].y;
-                therow.at<float>(0,1)=shape1[contPt].x;
+                therow.at<float>(0,1)=-shape1[contPt].x;
                 therow.at<float>(0,3)=1;
                 therow.row(0).copyTo(matM.row(ii));
                 matP.at<float>(ii,0) = shape2[contPt].y;

modules/ts/src/ts.cpp

@@ -921,6 +921,14 @@ inline static void recordPropertyVerbose(const std::string & property,
     }
 }
 
+inline static void recordPropertyVerbose(const std::string& property, const std::string& msg,
+                                         const char* value, const char* build_value = NULL)
+{
+    return recordPropertyVerbose(property, msg,
+        value ? std::string(value) : std::string(),
+        build_value ? std::string(build_value) : std::string());
+}
+
 #ifdef _DEBUG
 #define CV_TEST_BUILD_CONFIG "Debug"
 #else

modules/videoio/src/cap_xine.cpp

@@ -107,7 +107,7 @@ class XINECapture : public IVideoCapture
 
     bool open(const char *filename)
     {
-        CV_Assert(!xine, !stream, !vo_port);
+        CV_Assert_N(!xine, !stream, !vo_port);
         char configfile[2048] = {0};
 
         xine = xine_new();
@@ -207,7 +207,7 @@ class XINECapture : public IVideoCapture
 
     double getProperty(int property_id) const CV_OVERRIDE
     {
-        CV_Assert(xine, vo_port, stream);
+        CV_Assert_N(xine, vo_port, stream);
         int pos_t, pos_l, length;
         bool res = (bool)xine_get_pos_length(stream, &pos_l, &pos_t, &length);
 
@@ -240,7 +240,7 @@ class XINECapture : public IVideoCapture
 protected:
     bool oldSeekFrame(int f)
     {
-        CV_Assert(xine, vo_port, stream);
+        CV_Assert_N(xine, vo_port, stream);
         // no need to seek if we are already there...
         if (f == frame_number)
         {
@@ -290,7 +290,7 @@ protected:
 
     bool seekFrame(int f)
     {
-        CV_Assert(xine, vo_port, stream);
+        CV_Assert_N(xine, vo_port, stream);
         if (seekable)
         {
             int new_time = (int)((f + 1) * (float)frame_duration);
@@ -309,7 +309,7 @@ protected:
 
     bool seekTime(int t)
     {
-        CV_Assert(xine, vo_port, stream);
+        CV_Assert_N(xine, vo_port, stream);
         if (seekable)
         {
             if (xine_play(stream, 0, t))
@@ -328,7 +328,7 @@ protected:
 
     bool seekRatio(double ratio)
     {
-        CV_Assert(xine, vo_port, stream);
+        CV_Assert_N(xine, vo_port, stream);
         if (ratio > 1 || ratio < 0)
             return false;
         if (seekable)

modules/viz/src/widget.cpp

@@ -301,6 +301,7 @@ void cv::viz::Widget3D::applyTransform(const Affine3d &transform)
     vtkSmartPointer<vtkPolyDataMapper> mapper = vtkPolyDataMapper::SafeDownCast(actor->GetMapper());
     CV_Assert("Widget doesn't have a polydata mapper" && mapper);
 
+    mapper->Update(); // #10945
     VtkUtils::SetInputData(mapper, VtkUtils::TransformPolydata(mapper->GetInput(), transform));
     mapper->Update();
 }

platforms/android/build_sdk.py

@@ -106,9 +106,13 @@ class ABI:
         self.cmake_vars = dict(
             ANDROID_STL="gnustl_static",
             ANDROID_ABI=self.name,
-            ANDROID_TOOLCHAIN_NAME=toolchain,
             ANDROID_PLATFORM_ID=platform_id,
         )
+        if toolchain is not None:
+            self.cmake_vars['ANDROID_TOOLCHAIN_NAME'] = toolchain
+        else:
+            self.cmake_vars['ANDROID_TOOLCHAIN'] = 'clang'
+            self.cmake_vars['ANDROID_STL'] = 'c++_static'
         if ndk_api_level:
             self.cmake_vars['ANDROID_NATIVE_API_LEVEL'] = ndk_api_level
         self.cmake_vars.update(cmake_vars)
@@ -206,7 +210,7 @@ class Builder:
         # Add extra data
         apkxmldest = check_dir(os.path.join(apkdest, "res", "xml"), create=True)
         apklibdest = check_dir(os.path.join(apkdest, "libs", abi.name), create=True)
-        for ver, d in self.extra_packs + [("3.4.2", os.path.join(self.libdest, "lib"))]:
+        for ver, d in self.extra_packs + [("3.4.3", os.path.join(self.libdest, "lib"))]:
             r = ET.Element("library", attrib={"version": ver})
             log.info("Adding libraries from %s", d)
 

platforms/android/ndk-17.config.py

@@ -0,0 +1,6 @@
+ABIs = [
+    ABI("2", "armeabi-v7a", None, cmake_vars=dict(ANDROID_ABI='armeabi-v7a with NEON')),
+    ABI("3", "arm64-v8a",   None),
+    ABI("5", "x86_64",      None),
+    ABI("4", "x86",         None),
+]

platforms/android/service/engine/AndroidManifest.xml

@@ -1,8 +1,8 @@
 <?xml version="1.0" encoding="utf-8"?>
 <manifest xmlns:android="http://schemas.android.com/apk/res/android"
     package="org.opencv.engine"
-    android:versionCode="342@ANDROID_PLATFORM_ID@"
+    android:versionCode="343@ANDROID_PLATFORM_ID@"
-    android:versionName="3.42">
+    android:versionName="3.43">
 
     <uses-sdk android:minSdkVersion="@ANDROID_NATIVE_API_LEVEL@" android:targetSdkVersion="22"/>
     <uses-feature android:name="android.hardware.touchscreen" android:required="false"/>

platforms/android/service/engine/src/org/opencv/engine/OpenCVEngineService.java

@@ -137,7 +137,7 @@ public class OpenCVEngineService extends Service {
 
             @Override
             public int getEngineVersion() throws RemoteException {
-                int version = 3420;
+                int version = 3430;
                 try {
                     version = getPackageManager().getPackageInfo(getPackageName(), 0).versionCode;
                 } catch (NameNotFoundException e) {

platforms/android/service/readme.txt

@@ -12,7 +12,7 @@ manually using adb tool:
 
     adb install <path-to-OpenCV-sdk>/apk/OpenCV_<version>_Manager_<app_version>_<platform>.apk
 
-Example: OpenCV_3.4.2-dev_Manager_3.42_armeabi-v7a.apk
+Example: OpenCV_3.4.3-dev_Manager_3.43_armeabi-v7a.apk
 
 Use the list of platforms below to determine proper OpenCV Manager package for your device:
 

platforms/maven/opencv-it/pom.xml

@@ -4,7 +4,7 @@
     <parent>
         <groupId>org.opencv</groupId>
         <artifactId>opencv-parent</artifactId>
-        <version>3.4.2</version>
+        <version>3.4.3</version>
     </parent>
     <groupId>org.opencv</groupId>
     <artifactId>opencv-it</artifactId>

platforms/maven/opencv/pom.xml

@@ -4,7 +4,7 @@
     <parent>
         <groupId>org.opencv</groupId>
         <artifactId>opencv-parent</artifactId>
-        <version>3.4.2</version>
+        <version>3.4.3</version>
     </parent>
     <groupId>org.opencv</groupId>
     <artifactId>opencv</artifactId>

platforms/maven/pom.xml

@@ -3,7 +3,7 @@
     <modelVersion>4.0.0</modelVersion>
     <groupId>org.opencv</groupId>
     <artifactId>opencv-parent</artifactId>
-    <version>3.4.2</version>
+    <version>3.4.3</version>
     <packaging>pom</packaging>
     <name>OpenCV Parent POM</name>
     <licenses>

samples/dnn/mask_rcnn.py

@@ -0,0 +1,143 @@
+import cv2 as cv
+import argparse
+import numpy as np
+
+parser = argparse.ArgumentParser(description=
+        'Use this script to run Mask-RCNN object detection and semantic '
+        'segmentation network from TensorFlow Object Detection API.')
+parser.add_argument('--input', help='Path to input image or video file. Skip this argument to capture frames from a camera.')
+parser.add_argument('--model', required=True, help='Path to a .pb file with weights.')
+parser.add_argument('--config', required=True, help='Path to a .pxtxt file contains network configuration.')
+parser.add_argument('--classes', help='Optional path to a text file with names of classes.')
+parser.add_argument('--colors', help='Optional path to a text file with colors for an every class. '
+                                     'An every color is represented with three values from 0 to 255 in BGR channels order.')
+parser.add_argument('--width', type=int, default=800,
+                    help='Preprocess input image by resizing to a specific width.')
+parser.add_argument('--height', type=int, default=800,
+                    help='Preprocess input image by resizing to a specific height.')
+parser.add_argument('--thr', type=float, default=0.5, help='Confidence threshold')
+args = parser.parse_args()
+
+np.random.seed(324)
+
+# Load names of classes
+classes = None
+if args.classes:
+    with open(args.classes, 'rt') as f:
+        classes = f.read().rstrip('\n').split('\n')
+
+# Load colors
+colors = None
+if args.colors:
+    with open(args.colors, 'rt') as f:
+        colors = [np.array(color.split(' '), np.uint8) for color in f.read().rstrip('\n').split('\n')]
+
+legend = None
+def showLegend(classes):
+    global legend
+    if not classes is None and legend is None:
+        blockHeight = 30
+        assert(len(classes) == len(colors))
+
+        legend = np.zeros((blockHeight * len(colors), 200, 3), np.uint8)
+        for i in range(len(classes)):
+            block = legend[i * blockHeight:(i + 1) * blockHeight]
+            block[:,:] = colors[i]
+            cv.putText(block, classes[i], (0, blockHeight/2), cv.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255))
+
+        cv.namedWindow('Legend', cv.WINDOW_NORMAL)
+        cv.imshow('Legend', legend)
+        classes = None
+
+
+def drawBox(frame, classId, conf, left, top, right, bottom):
+    # Draw a bounding box.
+    cv.rectangle(frame, (left, top), (right, bottom), (0, 255, 0))
+
+    label = '%.2f' % conf
+
+    # Print a label of class.
+    if classes:
+        assert(classId < len(classes))
+        label = '%s: %s' % (classes[classId], label)
+
+    labelSize, baseLine = cv.getTextSize(label, cv.FONT_HERSHEY_SIMPLEX, 0.5, 1)
+    top = max(top, labelSize[1])
+    cv.rectangle(frame, (left, top - labelSize[1]), (left + labelSize[0], top + baseLine), (255, 255, 255), cv.FILLED)
+    cv.putText(frame, label, (left, top), cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0))
+
+
+# Load a network
+net = cv.dnn.readNet(args.model, args.config)
+net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
+
+winName = 'Mask-RCNN in OpenCV'
+cv.namedWindow(winName, cv.WINDOW_NORMAL)
+
+cap = cv.VideoCapture(args.input if args.input else 0)
+legend = None
+while cv.waitKey(1) < 0:
+    hasFrame, frame = cap.read()
+    if not hasFrame:
+        cv.waitKey()
+        break
+
+    frameH = frame.shape[0]
+    frameW = frame.shape[1]
+
+    # Create a 4D blob from a frame.
+    blob = cv.dnn.blobFromImage(frame, size=(args.width, args.height), swapRB=True, crop=False)
+
+    # Run a model
+    net.setInput(blob)
+
+    boxes, masks = net.forward(['detection_out_final', 'detection_masks'])
+
+    numClasses = masks.shape[1]
+    numDetections = boxes.shape[2]
+
+    # Draw segmentation
+    if not colors:
+        # Generate colors
+        colors = [np.array([0, 0, 0], np.uint8)]
+        for i in range(1, numClasses + 1):
+            colors.append((colors[i - 1] + np.random.randint(0, 256, [3], np.uint8)) / 2)
+        del colors[0]
+
+    boxesToDraw = []
+    for i in range(numDetections):
+        box = boxes[0, 0, i]
+        mask = masks[i]
+        score = box[2]
+        if score > args.thr:
+            classId = int(box[1])
+            left = int(frameW * box[3])
+            top = int(frameH * box[4])
+            right = int(frameW * box[5])
+            bottom = int(frameH * box[6])
+
+            left = max(0, min(left, frameW - 1))
+            top = max(0, min(top, frameH - 1))
+            right = max(0, min(right, frameW - 1))
+            bottom = max(0, min(bottom, frameH - 1))
+
+            boxesToDraw.append([frame, classId, score, left, top, right, bottom])
+
+            classMask = mask[classId]
+            classMask = cv.resize(classMask, (right - left + 1, bottom - top + 1))
+            mask = (classMask > 0.5)
+
+            roi = frame[top:bottom+1, left:right+1][mask]
+            frame[top:bottom+1, left:right+1][mask] = (0.7 * colors[classId] + 0.3 * roi).astype(np.uint8)
+
+    for box in boxesToDraw:
+        drawBox(*box)
+
+    # Put efficiency information.
+    t, _ = net.getPerfProfile()
+    label = 'Inference time: %.2f ms' % (t * 1000.0 / cv.getTickFrequency())
+    cv.putText(frame, label, (0, 15), cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0))
+
+    showLegend(classes)
+
+    cv.imshow(winName, frame)

samples/dnn/tf_text_graph_common.py

@@ -23,3 +23,98 @@ def addConstNode(name, values, graph_def):
     node.op = 'Const'
     text_format.Merge(tensorMsg(values), node.attr["value"])
     graph_def.node.extend([node])
+
+
+def addSlice(inp, out, begins, sizes, graph_def):
+    beginsNode = NodeDef()
+    beginsNode.name = out + '/begins'
+    beginsNode.op = 'Const'
+    text_format.Merge(tensorMsg(begins), beginsNode.attr["value"])
+    graph_def.node.extend([beginsNode])
+
+    sizesNode = NodeDef()
+    sizesNode.name = out + '/sizes'
+    sizesNode.op = 'Const'
+    text_format.Merge(tensorMsg(sizes), sizesNode.attr["value"])
+    graph_def.node.extend([sizesNode])
+
+    sliced = NodeDef()
+    sliced.name = out
+    sliced.op = 'Slice'
+    sliced.input.append(inp)
+    sliced.input.append(beginsNode.name)
+    sliced.input.append(sizesNode.name)
+    graph_def.node.extend([sliced])
+
+
+def addReshape(inp, out, shape, graph_def):
+    shapeNode = NodeDef()
+    shapeNode.name = out + '/shape'
+    shapeNode.op = 'Const'
+    text_format.Merge(tensorMsg(shape), shapeNode.attr["value"])
+    graph_def.node.extend([shapeNode])
+
+    reshape = NodeDef()
+    reshape.name = out
+    reshape.op = 'Reshape'
+    reshape.input.append(inp)
+    reshape.input.append(shapeNode.name)
+    graph_def.node.extend([reshape])
+
+
+def addSoftMax(inp, out, graph_def):
+    softmax = NodeDef()
+    softmax.name = out
+    softmax.op = 'Softmax'
+    text_format.Merge('i: -1', softmax.attr['axis'])
+    softmax.input.append(inp)
+    graph_def.node.extend([softmax])
+
+
+def addFlatten(inp, out, graph_def):
+    flatten = NodeDef()
+    flatten.name = out
+    flatten.op = 'Flatten'
+    flatten.input.append(inp)
+    graph_def.node.extend([flatten])
+
+
+# Removes Identity nodes
+def removeIdentity(graph_def):
+    identities = {}
+    for node in graph_def.node:
+        if node.op == 'Identity':
+            identities[node.name] = node.input[0]
+            graph_def.node.remove(node)
+
+    for node in graph_def.node:
+        for i in range(len(node.input)):
+            if node.input[i] in identities:
+                node.input[i] = identities[node.input[i]]
+
+
+def removeUnusedNodesAndAttrs(to_remove, graph_def):
+    unusedAttrs = ['T', 'Tshape', 'N', 'Tidx', 'Tdim', 'use_cudnn_on_gpu',
+                   'Index', 'Tperm', 'is_training', 'Tpaddings']
+
+    removedNodes = []
+
+    for i in reversed(range(len(graph_def.node))):
+        op = graph_def.node[i].op
+        name = graph_def.node[i].name
+
+        if op == 'Const' or to_remove(name, op):
+            if op != 'Const':
+                removedNodes.append(name)
+
+            del graph_def.node[i]
+        else:
+            for attr in unusedAttrs:
+                if attr in graph_def.node[i].attr:
+                    del graph_def.node[i].attr[attr]
+
+    # Remove references to removed nodes except Const nodes.
+    for node in graph_def.node:
+        for i in reversed(range(len(node.input))):
+            if node.input[i] in removedNodes:
+                del node.input[i]

samples/dnn/tf_text_graph_faster_rcnn.py

@@ -6,7 +6,7 @@ from tensorflow.core.framework.node_def_pb2 import NodeDef
 from tensorflow.tools.graph_transforms import TransformGraph
 from google.protobuf import text_format
 
-from tf_text_graph_common import tensorMsg, addConstNode
+from tf_text_graph_common import *
 
 parser = argparse.ArgumentParser(description='Run this script to get a text graph of '
                                              'SSD model from TensorFlow Object Detection API. '
@@ -37,50 +37,17 @@ scopesToIgnore = ('FirstStageFeatureExtractor/Assert',
                   'FirstStageFeatureExtractor/GreaterEqual',
                   'FirstStageFeatureExtractor/LogicalAnd')
 
-unusedAttrs = ['T', 'Tshape', 'N', 'Tidx', 'Tdim', 'use_cudnn_on_gpu',
-               'Index', 'Tperm', 'is_training', 'Tpaddings']
-
 # Read the graph.
 with tf.gfile.FastGFile(args.input, 'rb') as f:
     graph_def = tf.GraphDef()
     graph_def.ParseFromString(f.read())
 
-# Removes Identity nodes
+removeIdentity(graph_def)
-def removeIdentity():
-    identities = {}
-    for node in graph_def.node:
-        if node.op == 'Identity':
-            identities[node.name] = node.input[0]
-            graph_def.node.remove(node)
 
-    for node in graph_def.node:
+def to_remove(name, op):
-        for i in range(len(node.input)):
+    return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep)
-            if node.input[i] in identities:
-                node.input[i] = identities[node.input[i]]
 
-removeIdentity()
+removeUnusedNodesAndAttrs(to_remove, graph_def)
-
-removedNodes = []
-
-for i in reversed(range(len(graph_def.node))):
-    op = graph_def.node[i].op
-    name = graph_def.node[i].name
-
-    if op == 'Const' or name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep):
-        if op != 'Const':
-            removedNodes.append(name)
-
-        del graph_def.node[i]
-    else:
-        for attr in unusedAttrs:
-            if attr in graph_def.node[i].attr:
-                del graph_def.node[i].attr[attr]
-
-# Remove references to removed nodes except Const nodes.
-for node in graph_def.node:
-    for i in reversed(range(len(node.input))):
-        if node.input[i] in removedNodes:
-            del node.input[i]
 
 
 # Connect input node to the first layer
@@ -95,68 +62,18 @@ while True:
     if node.op == 'CropAndResize':
         break
 
-def addSlice(inp, out, begins, sizes):
-    beginsNode = NodeDef()
-    beginsNode.name = out + '/begins'
-    beginsNode.op = 'Const'
-    text_format.Merge(tensorMsg(begins), beginsNode.attr["value"])
-    graph_def.node.extend([beginsNode])
-
-    sizesNode = NodeDef()
-    sizesNode.name = out + '/sizes'
-    sizesNode.op = 'Const'
-    text_format.Merge(tensorMsg(sizes), sizesNode.attr["value"])
-    graph_def.node.extend([sizesNode])
-
-    sliced = NodeDef()
-    sliced.name = out
-    sliced.op = 'Slice'
-    sliced.input.append(inp)
-    sliced.input.append(beginsNode.name)
-    sliced.input.append(sizesNode.name)
-    graph_def.node.extend([sliced])
-
-def addReshape(inp, out, shape):
-    shapeNode = NodeDef()
-    shapeNode.name = out + '/shape'
-    shapeNode.op = 'Const'
-    text_format.Merge(tensorMsg(shape), shapeNode.attr["value"])
-    graph_def.node.extend([shapeNode])
-
-    reshape = NodeDef()
-    reshape.name = out
-    reshape.op = 'Reshape'
-    reshape.input.append(inp)
-    reshape.input.append(shapeNode.name)
-    graph_def.node.extend([reshape])
-
-def addSoftMax(inp, out):
-    softmax = NodeDef()
-    softmax.name = out
-    softmax.op = 'Softmax'
-    text_format.Merge('i: -1', softmax.attr['axis'])
-    softmax.input.append(inp)
-    graph_def.node.extend([softmax])
-
-def addFlatten(inp, out):
-    flatten = NodeDef()
-    flatten.name = out
-    flatten.op = 'Flatten'
-    flatten.input.append(inp)
-    graph_def.node.extend([flatten])
-
 addReshape('FirstStageBoxPredictor/ClassPredictor/BiasAdd',
-           'FirstStageBoxPredictor/ClassPredictor/reshape_1', [0, -1, 2])
+           'FirstStageBoxPredictor/ClassPredictor/reshape_1', [0, -1, 2], graph_def)
 
 addSoftMax('FirstStageBoxPredictor/ClassPredictor/reshape_1',
-           'FirstStageBoxPredictor/ClassPredictor/softmax')  # Compare with Reshape_4
+           'FirstStageBoxPredictor/ClassPredictor/softmax', graph_def)  # Compare with Reshape_4
 
 addFlatten('FirstStageBoxPredictor/ClassPredictor/softmax',
-           'FirstStageBoxPredictor/ClassPredictor/softmax/flatten')
+           'FirstStageBoxPredictor/ClassPredictor/softmax/flatten', graph_def)
 
 # Compare with FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd
 addFlatten('FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd',
-           'FirstStageBoxPredictor/BoxEncodingPredictor/flatten')
+           'FirstStageBoxPredictor/BoxEncodingPredictor/flatten', graph_def)
 
 proposals = NodeDef()
 proposals.name = 'proposals'  # Compare with ClipToWindow/Gather/Gather (NOTE: normalized)
@@ -218,14 +135,14 @@ graph_def.node.extend([clipByValueNode])
 for node in reversed(topNodes):
     graph_def.node.extend([node])
 
-addSoftMax('SecondStageBoxPredictor/Reshape_1', 'SecondStageBoxPredictor/Reshape_1/softmax')
+addSoftMax('SecondStageBoxPredictor/Reshape_1', 'SecondStageBoxPredictor/Reshape_1/softmax', graph_def)
 
 addSlice('SecondStageBoxPredictor/Reshape_1/softmax',
          'SecondStageBoxPredictor/Reshape_1/slice',
-         [0, 0, 1], [-1, -1, -1])
+         [0, 0, 1], [-1, -1, -1], graph_def)
 
 addReshape('SecondStageBoxPredictor/Reshape_1/slice',
-          'SecondStageBoxPredictor/Reshape_1/Reshape', [1, -1])
+          'SecondStageBoxPredictor/Reshape_1/Reshape', [1, -1], graph_def)
 
 # Replace Flatten subgraph onto a single node.
 for i in reversed(range(len(graph_def.node))):
@@ -255,7 +172,7 @@ for node in graph_def.node:
 ################################################################################
 ### Postprocessing
 ################################################################################
-addSlice('detection_out/clip_by_value', 'detection_out/slice', [0, 0, 0, 3], [-1, -1, -1, 4])
+addSlice('detection_out/clip_by_value', 'detection_out/slice', [0, 0, 0, 3], [-1, -1, -1, 4], graph_def)
 
 variance = NodeDef()
 variance.name = 'proposals/variance'
@@ -271,8 +188,8 @@ varianceEncoder.input.append(variance.name)
 text_format.Merge('i: 2', varianceEncoder.attr["axis"])
 graph_def.node.extend([varianceEncoder])
 
-addReshape('detection_out/slice', 'detection_out/slice/reshape', [1, 1, -1])
+addReshape('detection_out/slice', 'detection_out/slice/reshape', [1, 1, -1], graph_def)
-addFlatten('variance_encoded', 'variance_encoded/flatten')
+addFlatten('variance_encoded', 'variance_encoded/flatten', graph_def)
 
 detectionOut = NodeDef()
 detectionOut.name = 'detection_out_final'

samples/dnn/tf_text_graph_mask_rcnn.py

@@ -0,0 +1,230 @@
+import argparse
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.core.framework.node_def_pb2 import NodeDef
+from tensorflow.tools.graph_transforms import TransformGraph
+from google.protobuf import text_format
+
+from tf_text_graph_common import *
+
+parser = argparse.ArgumentParser(description='Run this script to get a text graph of '
+                                             'Mask-RCNN model from TensorFlow Object Detection API. '
+                                             'Then pass it with .pb file to cv::dnn::readNetFromTensorflow function.')
+parser.add_argument('--input', required=True, help='Path to frozen TensorFlow graph.')
+parser.add_argument('--output', required=True, help='Path to output text graph.')
+parser.add_argument('--num_classes', default=90, type=int, help='Number of trained classes.')
+parser.add_argument('--scales', default=[0.25, 0.5, 1.0, 2.0], type=float, nargs='+',
+                    help='Hyper-parameter of grid_anchor_generator from a config file.')
+parser.add_argument('--aspect_ratios', default=[0.5, 1.0, 2.0], type=float, nargs='+',
+                    help='Hyper-parameter of grid_anchor_generator from a config file.')
+parser.add_argument('--features_stride', default=16, type=float, nargs='+',
+                    help='Hyper-parameter from a config file.')
+args = parser.parse_args()
+
+scopesToKeep = ('FirstStageFeatureExtractor', 'Conv',
+                'FirstStageBoxPredictor/BoxEncodingPredictor',
+                'FirstStageBoxPredictor/ClassPredictor',
+                'CropAndResize',
+                'MaxPool2D',
+                'SecondStageFeatureExtractor',
+                'SecondStageBoxPredictor',
+                'Preprocessor/sub',
+                'Preprocessor/mul',
+                'image_tensor')
+
+scopesToIgnore = ('FirstStageFeatureExtractor/Assert',
+                  'FirstStageFeatureExtractor/Shape',
+                  'FirstStageFeatureExtractor/strided_slice',
+                  'FirstStageFeatureExtractor/GreaterEqual',
+                  'FirstStageFeatureExtractor/LogicalAnd')
+
+
+# Read the graph.
+with tf.gfile.FastGFile(args.input, 'rb') as f:
+    graph_def = tf.GraphDef()
+    graph_def.ParseFromString(f.read())
+
+removeIdentity(graph_def)
+
+def to_remove(name, op):
+    return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep)
+
+removeUnusedNodesAndAttrs(to_remove, graph_def)
+
+
+# Connect input node to the first layer
+assert(graph_def.node[0].op == 'Placeholder')
+graph_def.node[1].input.insert(0, graph_def.node[0].name)
+
+# Temporarily remove top nodes.
+topNodes = []
+numCropAndResize = 0
+while True:
+    node = graph_def.node.pop()
+    topNodes.append(node)
+    if node.op == 'CropAndResize':
+        numCropAndResize += 1
+        if numCropAndResize == 2:
+            break
+
+addReshape('FirstStageBoxPredictor/ClassPredictor/BiasAdd',
+           'FirstStageBoxPredictor/ClassPredictor/reshape_1', [0, -1, 2], graph_def)
+
+addSoftMax('FirstStageBoxPredictor/ClassPredictor/reshape_1',
+           'FirstStageBoxPredictor/ClassPredictor/softmax', graph_def)  # Compare with Reshape_4
+
+addFlatten('FirstStageBoxPredictor/ClassPredictor/softmax',
+           'FirstStageBoxPredictor/ClassPredictor/softmax/flatten', graph_def)
+
+# Compare with FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd
+addFlatten('FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd',
+           'FirstStageBoxPredictor/BoxEncodingPredictor/flatten', graph_def)
+
+proposals = NodeDef()
+proposals.name = 'proposals'  # Compare with ClipToWindow/Gather/Gather (NOTE: normalized)
+proposals.op = 'PriorBox'
+proposals.input.append('FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd')
+proposals.input.append(graph_def.node[0].name)  # image_tensor
+
+text_format.Merge('b: false', proposals.attr["flip"])
+text_format.Merge('b: true', proposals.attr["clip"])
+text_format.Merge('f: %f' % args.features_stride, proposals.attr["step"])
+text_format.Merge('f: 0.0', proposals.attr["offset"])
+text_format.Merge(tensorMsg([0.1, 0.1, 0.2, 0.2]), proposals.attr["variance"])
+
+widths = []
+heights = []
+for a in args.aspect_ratios:
+    for s in args.scales:
+        ar = np.sqrt(a)
+        heights.append((args.features_stride**2) * s / ar)
+        widths.append((args.features_stride**2) * s * ar)
+
+text_format.Merge(tensorMsg(widths), proposals.attr["width"])
+text_format.Merge(tensorMsg(heights), proposals.attr["height"])
+
+graph_def.node.extend([proposals])
+
+# Compare with Reshape_5
+detectionOut = NodeDef()
+detectionOut.name = 'detection_out'
+detectionOut.op = 'DetectionOutput'
+
+detectionOut.input.append('FirstStageBoxPredictor/BoxEncodingPredictor/flatten')
+detectionOut.input.append('FirstStageBoxPredictor/ClassPredictor/softmax/flatten')
+detectionOut.input.append('proposals')
+
+text_format.Merge('i: 2', detectionOut.attr['num_classes'])
+text_format.Merge('b: true', detectionOut.attr['share_location'])
+text_format.Merge('i: 0', detectionOut.attr['background_label_id'])
+text_format.Merge('f: 0.7', detectionOut.attr['nms_threshold'])
+text_format.Merge('i: 6000', detectionOut.attr['top_k'])
+text_format.Merge('s: "CENTER_SIZE"', detectionOut.attr['code_type'])
+text_format.Merge('i: 100', detectionOut.attr['keep_top_k'])
+text_format.Merge('b: true', detectionOut.attr['clip'])
+
+graph_def.node.extend([detectionOut])
+
+# Save as text.
+for node in reversed(topNodes):
+    if node.op != 'CropAndResize':
+        graph_def.node.extend([node])
+        topNodes.pop()
+    else:
+        if numCropAndResize == 1:
+            break
+        else:
+            graph_def.node.extend([node])
+            topNodes.pop()
+            numCropAndResize -= 1
+
+addSoftMax('SecondStageBoxPredictor/Reshape_1', 'SecondStageBoxPredictor/Reshape_1/softmax', graph_def)
+
+addSlice('SecondStageBoxPredictor/Reshape_1/softmax',
+         'SecondStageBoxPredictor/Reshape_1/slice',
+         [0, 0, 1], [-1, -1, -1], graph_def)
+
+addReshape('SecondStageBoxPredictor/Reshape_1/slice',
+          'SecondStageBoxPredictor/Reshape_1/Reshape', [1, -1], graph_def)
+
+# Replace Flatten subgraph onto a single node.
+for i in reversed(range(len(graph_def.node))):
+    if graph_def.node[i].op == 'CropAndResize':
+        graph_def.node[i].input.insert(1, 'detection_out')
+
+    if graph_def.node[i].name == 'SecondStageBoxPredictor/Reshape':
+        addConstNode('SecondStageBoxPredictor/Reshape/shape2', [1, -1, 4], graph_def)
+
+        graph_def.node[i].input.pop()
+        graph_def.node[i].input.append('SecondStageBoxPredictor/Reshape/shape2')
+
+    if graph_def.node[i].name in ['SecondStageBoxPredictor/Flatten/flatten/Shape',
+                                  'SecondStageBoxPredictor/Flatten/flatten/strided_slice',
+                                  'SecondStageBoxPredictor/Flatten/flatten/Reshape/shape']:
+        del graph_def.node[i]
+
+for node in graph_def.node:
+    if node.name == 'SecondStageBoxPredictor/Flatten/flatten/Reshape':
+        node.op = 'Flatten'
+        node.input.pop()
+
+    if node.name in ['FirstStageBoxPredictor/BoxEncodingPredictor/Conv2D',
+                     'SecondStageBoxPredictor/BoxEncodingPredictor/MatMul']:
+        text_format.Merge('b: true', node.attr["loc_pred_transposed"])
+
+################################################################################
+### Postprocessing
+################################################################################
+addSlice('detection_out', 'detection_out/slice', [0, 0, 0, 3], [-1, -1, -1, 4], graph_def)
+
+variance = NodeDef()
+variance.name = 'proposals/variance'
+variance.op = 'Const'
+text_format.Merge(tensorMsg([0.1, 0.1, 0.2, 0.2]), variance.attr["value"])
+graph_def.node.extend([variance])
+
+varianceEncoder = NodeDef()
+varianceEncoder.name = 'variance_encoded'
+varianceEncoder.op = 'Mul'
+varianceEncoder.input.append('SecondStageBoxPredictor/Reshape')
+varianceEncoder.input.append(variance.name)
+text_format.Merge('i: 2', varianceEncoder.attr["axis"])
+graph_def.node.extend([varianceEncoder])
+
+addReshape('detection_out/slice', 'detection_out/slice/reshape', [1, 1, -1], graph_def)
+addFlatten('variance_encoded', 'variance_encoded/flatten', graph_def)
+
+detectionOut = NodeDef()
+detectionOut.name = 'detection_out_final'
+detectionOut.op = 'DetectionOutput'
+
+detectionOut.input.append('variance_encoded/flatten')
+detectionOut.input.append('SecondStageBoxPredictor/Reshape_1/Reshape')
+detectionOut.input.append('detection_out/slice/reshape')
+
+text_format.Merge('i: %d' % args.num_classes, detectionOut.attr['num_classes'])
+text_format.Merge('b: false', detectionOut.attr['share_location'])
+text_format.Merge('i: %d' % (args.num_classes + 1), detectionOut.attr['background_label_id'])
+text_format.Merge('f: 0.6', detectionOut.attr['nms_threshold'])
+text_format.Merge('s: "CENTER_SIZE"', detectionOut.attr['code_type'])
+text_format.Merge('i: 100', detectionOut.attr['keep_top_k'])
+text_format.Merge('b: true', detectionOut.attr['clip'])
+text_format.Merge('b: true', detectionOut.attr['variance_encoded_in_target'])
+text_format.Merge('f: 0.3', detectionOut.attr['confidence_threshold'])
+text_format.Merge('b: false', detectionOut.attr['group_by_classes'])
+graph_def.node.extend([detectionOut])
+
+for node in reversed(topNodes):
+    graph_def.node.extend([node])
+
+for i in reversed(range(len(graph_def.node))):
+    if graph_def.node[i].op == 'CropAndResize':
+        graph_def.node[i].input.insert(1, 'detection_out_final')
+        break
+
+graph_def.node[-1].name = 'detection_masks'
+graph_def.node[-1].op = 'Sigmoid'
+graph_def.node[-1].input.pop()
+
+tf.train.write_graph(graph_def, "", args.output, as_text=True)

samples/dnn/tf_text_graph_ssd.py

@@ -15,7 +15,7 @@ from math import sqrt
 from tensorflow.core.framework.node_def_pb2 import NodeDef
 from tensorflow.tools.graph_transforms import TransformGraph
 from google.protobuf import text_format
-from tf_text_graph_common import tensorMsg, addConstNode
+from tf_text_graph_common import *
 
 parser = argparse.ArgumentParser(description='Run this script to get a text graph of '
                                              'SSD model from TensorFlow Object Detection API. '
@@ -41,10 +41,6 @@ args = parser.parse_args()
 keepOps = ['Conv2D', 'BiasAdd', 'Add', 'Relu6', 'Placeholder', 'FusedBatchNorm',
            'DepthwiseConv2dNative', 'ConcatV2', 'Mul', 'MaxPool', 'AvgPool', 'Identity']
 
-# Nodes attributes that could be removed because they are not used during import.
-unusedAttrs = ['T', 'data_format', 'Tshape', 'N', 'Tidx', 'Tdim', 'use_cudnn_on_gpu',
-               'Index', 'Tperm', 'is_training', 'Tpaddings']
-
 # Node with which prefixes should be removed
 prefixesToRemove = ('MultipleGridAnchorGenerator/', 'Postprocessor/', 'Preprocessor/')
 
@@ -66,7 +62,6 @@ def getUnconnectedNodes():
                 unconnected.remove(inp)
     return unconnected
 
-removedNodes = []
 
 # Detect unfused batch normalization nodes and fuse them.
 def fuse_batch_normalization():
@@ -118,41 +113,13 @@ def fuse_batch_normalization():
 
 fuse_batch_normalization()
 
-# Removes Identity nodes
+removeIdentity(graph_def)
-def removeIdentity():
-    identities = {}
-    for node in graph_def.node:
-        if node.op == 'Identity':
-            identities[node.name] = node.input[0]
-            graph_def.node.remove(node)
 
-    for node in graph_def.node:
+def to_remove(name, op):
-        for i in range(len(node.input)):
+    return (not op in keepOps) or name.startswith(prefixesToRemove)
-            if node.input[i] in identities:
-                node.input[i] = identities[node.input[i]]
 
-removeIdentity()
+removeUnusedNodesAndAttrs(to_remove, graph_def)
 
-# Remove extra nodes and attributes.
-for i in reversed(range(len(graph_def.node))):
-    op = graph_def.node[i].op
-    name = graph_def.node[i].name
-
-    if (not op in keepOps) or name.startswith(prefixesToRemove):
-        if op != 'Const':
-            removedNodes.append(name)
-
-        del graph_def.node[i]
-    else:
-        for attr in unusedAttrs:
-            if attr in graph_def.node[i].attr:
-                del graph_def.node[i].attr[attr]
-
-# Remove references to removed nodes except Const nodes.
-for node in graph_def.node:
-    for i in reversed(range(len(node.input))):
-        if node.input[i] in removedNodes:
-            del node.input[i]
 
 # Connect input node to the first layer
 assert(graph_def.node[0].op == 'Placeholder')
@@ -175,8 +142,8 @@ def addConcatNode(name, inputs, axisNodeName):
     concat.input.append(axisNodeName)
     graph_def.node.extend([concat])
 
-addConstNode('concat/axis_flatten', [-1])
+addConstNode('concat/axis_flatten', [-1], graph_def)
-addConstNode('PriorBox/concat/axis', [-2])
+addConstNode('PriorBox/concat/axis', [-2], graph_def)
 
 for label in ['ClassPredictor', 'BoxEncodingPredictor' if args.box_predictor is 'convolutional' else 'BoxPredictor']:
     concatInputs = []