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

doc/opencv.bib

@@ -620,7 +620,7 @@ @inproceedings{Lienhart02
   volume = {1},
   publisher = {IEEE}
 }
-@article{Lowe:2004:DIF:993451.996342,
+@article{Lowe04,
   author = {Lowe, David G.},
   title = {Distinctive Image Features from Scale-Invariant Keypoints},
   journal = {Int. J. Comput. Vision},

doc/py_tutorials/py_feature2d/py_feature_homography/py_feature_homography.markdown

@@ -44,7 +44,7 @@ img1 = cv.imread('box.png',0)          # queryImage
 img2 = cv.imread('box_in_scene.png',0) # trainImage
 
 # Initiate SIFT detector
-sift = cv.xfeatures2d.SIFT_create()
+sift = cv.SIFT_create()
 
 # find the keypoints and descriptors with SIFT
 kp1, des1 = sift.detectAndCompute(img1,None)

doc/py_tutorials/py_feature2d/py_matcher/py_matcher.markdown

@@ -110,7 +110,7 @@ img1 = cv.imread('box.png',cv.IMREAD_GRAYSCALE)          # queryImage
 img2 = cv.imread('box_in_scene.png',cv.IMREAD_GRAYSCALE) # trainImage
 
 # Initiate SIFT detector
-sift = cv.xfeatures2d.SIFT_create()
+sift = cv.SIFT_create()
 
 # find the keypoints and descriptors with SIFT
 kp1, des1 = sift.detectAndCompute(img1,None)
@@ -174,7 +174,7 @@ img1 = cv.imread('box.png',cv.IMREAD_GRAYSCALE)          # queryImage
 img2 = cv.imread('box_in_scene.png',cv.IMREAD_GRAYSCALE) # trainImage
 
 # Initiate SIFT detector
-sift = cv.xfeatures2d.SIFT_create()
+sift = cv.SIFT_create()
 
 # find the keypoints and descriptors with SIFT
 kp1, des1 = sift.detectAndCompute(img1,None)

doc/py_tutorials/py_feature2d/py_sift_intro/py_sift_intro.markdown

@@ -119,7 +119,7 @@ import cv2 as cv
 img = cv.imread('home.jpg')
 gray= cv.cvtColor(img,cv.COLOR_BGR2GRAY)
 
-sift = cv.xfeatures2d.SIFT_create()
+sift = cv.SIFT_create()
 kp = sift.detect(gray,None)
 
 img=cv.drawKeypoints(gray,kp,img)
@@ -151,7 +151,7 @@ Now to calculate the descriptor, OpenCV provides two methods.
 
 We will see the second method:
 @code{.py}
-sift = cv.xfeatures2d.SIFT_create()
+sift = cv.SIFT_create()
 kp, des = sift.detectAndCompute(gray,None)
 @endcode
 Here kp will be a list of keypoints and des is a numpy array of shape

doc/tutorials/features2d/feature_flann_matcher/feature_flann_matcher.markdown

@@ -27,7 +27,7 @@ Binary descriptors (ORB, BRISK, ...) are matched using the <a href="https://en.w
 This distance is equivalent to count the number of different elements for binary strings (population count after applying a XOR operation):
 \f[ d_{hamming} \left ( a,b \right ) = \sum_{i=0}^{n-1} \left ( a_i \oplus b_i \right ) \f]
 
-To filter the matches, Lowe proposed in @cite Lowe:2004:DIF:993451.996342 to use a distance ratio test to try to eliminate false matches.
+To filter the matches, Lowe proposed in @cite Lowe04 to use a distance ratio test to try to eliminate false matches.
 The distance ratio between the two nearest matches of a considered keypoint is computed and it is a good match when this value is below
 a threshold. Indeed, this ratio allows helping to discriminate between ambiguous matches (distance ratio between the two nearest neighbors
 is close to one) and well discriminated matches. The figure below from the SIFT paper illustrates the probability that a match is correct

modules/core/include/opencv2/core/hal/intrin_avx.hpp

@@ -159,6 +159,7 @@ struct v_uint8x32
             (char)v22, (char)v23, (char)v24, (char)v25, (char)v26, (char)v27,
             (char)v28, (char)v29, (char)v30, (char)v31);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint8x32() {}
 
     uchar get0() const { return (uchar)_v_cvtsi256_si32(val); }
@@ -184,6 +185,7 @@ struct v_int8x32
             v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20,
             v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_int8x32() {}
 
     schar get0() const { return (schar)_v_cvtsi256_si32(val); }
@@ -205,6 +207,7 @@ struct v_uint16x16
             (short)v4,  (short)v5,  (short)v6,  (short)v7,  (short)v8,  (short)v9,
             (short)v10, (short)v11, (short)v12, (short)v13, (short)v14, (short)v15);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint16x16() {}
 
     ushort get0() const { return (ushort)_v_cvtsi256_si32(val); }
@@ -225,6 +228,7 @@ struct v_int16x16
         val = _mm256_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7,
             v8, v9, v10, v11, v12, v13, v14, v15);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_int16x16() {}
 
     short get0() const { return (short)_v_cvtsi256_si32(val); }
@@ -243,6 +247,7 @@ struct v_uint32x8
         val = _mm256_setr_epi32((unsigned)v0, (unsigned)v1, (unsigned)v2,
             (unsigned)v3, (unsigned)v4, (unsigned)v5, (unsigned)v6, (unsigned)v7);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint32x8() {}
 
     unsigned get0() const { return (unsigned)_v_cvtsi256_si32(val); }
@@ -260,6 +265,7 @@ struct v_int32x8
     {
         val = _mm256_setr_epi32(v0, v1, v2, v3, v4, v5, v6, v7);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_int32x8() {}
 
     int get0() const { return _v_cvtsi256_si32(val); }
@@ -277,6 +283,7 @@ struct v_float32x8
     {
         val = _mm256_setr_ps(v0, v1, v2, v3, v4, v5, v6, v7);
     }
+    /* coverity[uninit_ctor]: suppress warning */
     v_float32x8() {}
 
     float get0() const { return _mm_cvtss_f32(_mm256_castps256_ps128(val)); }
@@ -291,7 +298,9 @@ struct v_uint64x4
     explicit v_uint64x4(__m256i v) : val(v) {}
     v_uint64x4(uint64 v0, uint64 v1, uint64 v2, uint64 v3)
     { val = _mm256_setr_epi64x((int64)v0, (int64)v1, (int64)v2, (int64)v3); }
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint64x4() {}
+
     uint64 get0() const
     {
     #if defined __x86_64__ || defined _M_X64
@@ -313,6 +322,7 @@ struct v_int64x4
     explicit v_int64x4(__m256i v) : val(v) {}
     v_int64x4(int64 v0, int64 v1, int64 v2, int64 v3)
     { val = _mm256_setr_epi64x(v0, v1, v2, v3); }
+    /* coverity[uninit_ctor]: suppress warning */
     v_int64x4() {}
 
     int64 get0() const
@@ -336,6 +346,7 @@ struct v_float64x4
     explicit v_float64x4(__m256d v) : val(v) {}
     v_float64x4(double v0, double v1, double v2, double v3)
     { val = _mm256_setr_pd(v0, v1, v2, v3); }
+    /* coverity[uninit_ctor]: suppress warning */
     v_float64x4() {}
 
     double get0() const { return _mm_cvtsd_f64(_mm256_castpd256_pd128(val)); }

modules/core/include/opencv2/core/hal/intrin_sse.hpp

@@ -75,6 +75,7 @@ struct v_uint8x16
     typedef __m128i vector_type;
     enum { nlanes = 16 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint8x16() {}
     explicit v_uint8x16(__m128i v) : val(v) {}
     v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
@@ -100,6 +101,7 @@ struct v_int8x16
     typedef __m128i vector_type;
     enum { nlanes = 16 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_int8x16() {}
     explicit v_int8x16(__m128i v) : val(v) {}
     v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
@@ -125,6 +127,7 @@ struct v_uint16x8
     typedef __m128i vector_type;
     enum { nlanes = 8 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint16x8() {}
     explicit v_uint16x8(__m128i v) : val(v) {}
     v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
@@ -147,6 +150,7 @@ struct v_int16x8
     typedef __m128i vector_type;
     enum { nlanes = 8 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_int16x8() {}
     explicit v_int16x8(__m128i v) : val(v) {}
     v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
@@ -169,6 +173,7 @@ struct v_uint32x4
     typedef __m128i vector_type;
     enum { nlanes = 4 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint32x4() {}
     explicit v_uint32x4(__m128i v) : val(v) {}
     v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
@@ -190,6 +195,7 @@ struct v_int32x4
     typedef __m128i vector_type;
     enum { nlanes = 4 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_int32x4() {}
     explicit v_int32x4(__m128i v) : val(v) {}
     v_int32x4(int v0, int v1, int v2, int v3)
@@ -211,6 +217,7 @@ struct v_float32x4
     typedef __m128 vector_type;
     enum { nlanes = 4 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_float32x4() {}
     explicit v_float32x4(__m128 v) : val(v) {}
     v_float32x4(float v0, float v1, float v2, float v3)
@@ -232,6 +239,7 @@ struct v_uint64x2
     typedef __m128i vector_type;
     enum { nlanes = 2 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_uint64x2() {}
     explicit v_uint64x2(__m128i v) : val(v) {}
     v_uint64x2(uint64 v0, uint64 v1)
@@ -259,6 +267,7 @@ struct v_int64x2
     typedef __m128i vector_type;
     enum { nlanes = 2 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_int64x2() {}
     explicit v_int64x2(__m128i v) : val(v) {}
     v_int64x2(int64 v0, int64 v1)
@@ -286,6 +295,7 @@ struct v_float64x2
     typedef __m128d vector_type;
     enum { nlanes = 2 };
 
+    /* coverity[uninit_ctor]: suppress warning */
     v_float64x2() {}
     explicit v_float64x2(__m128d v) : val(v) {}
     v_float64x2(double v0, double v1)

modules/dnn/src/dnn.cpp

@@ -2230,7 +2230,7 @@ struct Net::Impl
 
                     auto ieInpNode = inputNodes[i].dynamicCast<InfEngineNgraphNode>();
                     CV_Assert(oid < ieInpNode->node->get_output_size());
-#if INF_ENGINE_VER_MAJOR_GT(2020020000)
+#if INF_ENGINE_VER_MAJOR_GT(2020030000)
                     inputNodes[i] = Ptr<BackendNode>(new InfEngineNgraphNode(ieInpNode->node->get_output_as_single_output_node(oid)));
 #else
                     inputNodes[i] = Ptr<BackendNode>(new InfEngineNgraphNode(ieInpNode->node->get_output_as_single_output_node(oid, false)));
@@ -3450,6 +3450,8 @@ Net Net::Impl::createNetworkFromModelOptimizer(InferenceEngine::CNNNetwork& ieNe
 {
     CV_TRACE_FUNCTION();
 
+    CV_TRACE_REGION("register_inputs");
+
     std::vector<String> inputsNames;
     std::vector<MatShape> inp_shapes;
     for (auto& it : ieNet.getInputsInfo())
@@ -3468,6 +3470,8 @@ Net Net::Impl::createNetworkFromModelOptimizer(InferenceEngine::CNNNetwork& ieNe
         cvNet.setInputShape(inputsNames[inp_id], inp_shapes[inp_id]);
     }
 
+    CV_TRACE_REGION_NEXT("backendNode");
+
     Ptr<BackendNode> backendNode;
 #ifdef HAVE_DNN_NGRAPH
     if (DNN_BACKEND_INFERENCE_ENGINE_NGRAPH == getInferenceEngineBackendTypeParam())
@@ -3489,8 +3493,25 @@ Net Net::Impl::createNetworkFromModelOptimizer(InferenceEngine::CNNNetwork& ieNe
 #endif
     }
 
+    CV_TRACE_REGION_NEXT("register_outputs");
+
+#ifdef HAVE_DNN_NGRAPH
+    auto ngraphFunction = ieNet.getFunction();
+#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2020_2)
+    std::list< std::shared_ptr<ngraph::Node> > ngraphOperations;
+#else
+    std::vector< std::shared_ptr<ngraph::Node> > ngraphOperations;
+#endif
+    if (ngraphFunction)
+    {
+        ngraphOperations = ngraphFunction->get_ops();
+    }
+#endif
+
     for (auto& it : ieNet.getOutputsInfo())
     {
+        CV_TRACE_REGION("output");
+
         LayerParams lp;
         int lid = cvNet.addLayer(it.first, "", lp);
 
@@ -3499,15 +3520,38 @@ Net Net::Impl::createNetworkFromModelOptimizer(InferenceEngine::CNNNetwork& ieNe
 #ifdef HAVE_DNN_NGRAPH
         if (DNN_BACKEND_INFERENCE_ENGINE_NGRAPH == getInferenceEngineBackendTypeParam())
         {
+            const auto& outputName = it.first;
             Ptr<Layer> cvLayer(new NgraphBackendLayer(ieNet));
+            cvLayer->name = outputName;
+            cvLayer->type = "_unknown_";
 
-            InferenceEngine::CNNLayerPtr ieLayer = ieNet.getLayerByName(it.first.c_str());
-            CV_Assert(ieLayer);
+            if (ngraphFunction)
+            {
+                CV_TRACE_REGION("ngraph_function");
+                bool found = false;
+                for (const auto& op : ngraphOperations)
+                {
+                    CV_Assert(op);
+                    if (op->get_friendly_name() == outputName)
+                    {
+                        const std::string typeName = op->get_type_info().name;
+                        cvLayer->type = typeName;
+                        found = true;
+                        break;
+                    }
+                }
+                if (!found)
+                    CV_LOG_WARNING(NULL, "DNN/IE: Can't determine output layer type: '" << outputName << "'");
+            }
+            else
+            {
+                CV_TRACE_REGION("legacy_cnn_layer");
+                InferenceEngine::CNNLayerPtr ieLayer = ieNet.getLayerByName(it.first.c_str());
+                CV_Assert(ieLayer);
 
-            cvLayer->name = it.first;
-            cvLayer->type = ieLayer->type;
+                cvLayer->type = ieLayer->type;
+            }
             ld.layerInstance = cvLayer;
-
             ld.backendNodes[DNN_BACKEND_INFERENCE_ENGINE_NGRAPH] = backendNode;
         }
         else
@@ -3532,6 +3576,9 @@ Net Net::Impl::createNetworkFromModelOptimizer(InferenceEngine::CNNNetwork& ieNe
         for (int i = 0; i < inputsNames.size(); ++i)
             cvNet.connect(0, i, lid, i);
     }
+
+    CV_TRACE_REGION_NEXT("finalize");
+
     cvNet.setPreferableBackend(getInferenceEngineBackendTypeParam());
 
     cvNet.impl->skipInfEngineInit = true;

modules/dnn/src/onnx/onnx_graph_simplifier.cpp

@@ -382,6 +382,36 @@ class ResizeSubgraph2 : public ExtractScalesSubgraph
     }
 };
 
+class BatchNormalizationSubgraph : public Subgraph
+{
+public:
+    BatchNormalizationSubgraph()
+    {
+        int input = addNodeToMatch("");
+        int data1 = addNodeToMatch("Constant");
+        int data2 = addNodeToMatch("Constant");
+        int data3 = addNodeToMatch("Constant");
+        int data4 = addNodeToMatch("Constant");
+        int shape1 = addNodeToMatch("Constant");
+        int reshape1 = addNodeToMatch("Reshape", data1, shape1);
+        int shape2 = addNodeToMatch("Constant");
+        int reshape2 = addNodeToMatch("Reshape", data2, shape2);
+        int shape3 = addNodeToMatch("Constant");
+        int reshape3 = addNodeToMatch("Reshape", data3, shape3);
+        int shape4 = addNodeToMatch("Constant");
+        int reshape4 = addNodeToMatch("Reshape", data4, shape4);
+        int sqrtNode = addNodeToMatch("Sqrt", reshape3);
+        int A = addNodeToMatch("Constant");
+        int divNode = addNodeToMatch("Div", A, sqrtNode);
+        int mul1 = addNodeToMatch("Mul", reshape1, divNode);
+        int mul2 = addNodeToMatch("Mul", reshape4, mul1);
+        int sub = addNodeToMatch("Sub", reshape2, mul2);
+        int mul3 = addNodeToMatch("Mul", input, mul1);
+        addNodeToMatch("Add", mul3, sub);
+        setFusedNode("BatchNormalization", input, data1, data2, data4 ,data3);
+    }
+};
+
 void simplifySubgraphs(opencv_onnx::GraphProto& net)
 {
     std::vector<Ptr<Subgraph> > subgraphs;
@@ -394,6 +424,7 @@ void simplifySubgraphs(opencv_onnx::GraphProto& net)
     subgraphs.push_back(makePtr<NormalizeSubgraph1>());
     subgraphs.push_back(makePtr<NormalizeSubgraph2>());
     subgraphs.push_back(makePtr<NormalizeSubgraph3>());
+    subgraphs.push_back(makePtr<BatchNormalizationSubgraph>());
 
     simplifySubgraphs(Ptr<ImportGraphWrapper>(new ONNXGraphWrapper(net)), subgraphs);
 }