Replace Slice layer to Crop in Faster-RCNN networks from Caffe

modules/dnn/src/layers/crop_layer.cpp

@@ -41,6 +41,7 @@
 //M*/
 
 #include "../precomp.hpp"
+#include "../op_inf_engine.hpp"
 #include "layers_common.hpp"
 
 namespace cv
@@ -64,6 +65,12 @@ public:
         }
     }
 
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+        return backendId == DNN_BACKEND_OPENCV ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE && crop_ranges.size() == 4;
+    }
+
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
                          const int requiredOutputs,
                          std::vector<MatShape> &outputs,
@@ -109,7 +116,11 @@ public:
                 offset_final[i] = offset[i - start_axis];
         }
 
-        crop_ranges.resize(dims, Range::all());
+        crop_ranges.resize(dims);
+        for (int i = 0; i < start_axis; i++)
+        {
+            crop_ranges[i] = Range(0, inpBlob.size[i]);
+        }
         for (int i = start_axis; i < dims; i++)
         {
             if (offset_final[i] < 0 || offset_final[i] + inpSzBlob.size[i] > inpBlob.size[i])
@@ -138,6 +149,38 @@ public:
         input(&crop_ranges[0]).copyTo(output);
     }
 
+    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
+    {
+#ifdef HAVE_INF_ENGINE
+        InferenceEngine::LayerParams lp;
+        lp.name = name;
+        lp.type = "Crop";
+        lp.precision = InferenceEngine::Precision::FP32;
+        std::shared_ptr<InferenceEngine::CropLayer> ieLayer(new InferenceEngine::CropLayer(lp));
+
+        CV_Assert(crop_ranges.size() == 4);
+
+        ieLayer->axis.push_back(0);  // batch
+        ieLayer->offset.push_back(crop_ranges[0].start);
+        ieLayer->dim.push_back(crop_ranges[0].end - crop_ranges[0].start);
+
+        ieLayer->axis.push_back(1);  // channels
+        ieLayer->offset.push_back(crop_ranges[1].start);
+        ieLayer->dim.push_back(crop_ranges[1].end - crop_ranges[1].start);
+
+        ieLayer->axis.push_back(3);  // height
+        ieLayer->offset.push_back(crop_ranges[2].start);
+        ieLayer->dim.push_back(crop_ranges[2].end - crop_ranges[2].start);
+
+        ieLayer->axis.push_back(2);  // width
+        ieLayer->offset.push_back(crop_ranges[3].start);
+        ieLayer->dim.push_back(crop_ranges[3].end - crop_ranges[3].start);
+
+        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
+#endif  // HAVE_INF_ENGINE
+        return Ptr<BackendNode>();
+    }
+
     std::vector<Range> crop_ranges;
 };
 

modules/dnn/src/layers/fully_connected_layer.cpp

@@ -350,17 +350,33 @@ public:
             inshape = shape(outerSize, innerSize);
             outshape = shape(outerSize, numOutput);
 
-            UMat srcMat, dstMat;
+            UMat srcMat, dstMat, srcMat_fp32, dstMat_fp32;
             srcMat = inputs[i].reshape(1, inshape.size(), &inshape[0]);
             dstMat = outputs[i].reshape(1, outshape.size(), &outshape[0]);
 
-            cv::gemm(srcMat, weights, 1, noArray(), 0, dstMat, GEMM_2_T);
+            if (use_half)
+            {
+                convertFp16(srcMat, srcMat_fp32);
+                convertFp16(dstMat, dstMat_fp32);
+            }
+            else
+            {
+                srcMat_fp32 = srcMat;
+                dstMat_fp32 = dstMat;
+            }
+
+            cv::gemm(srcMat_fp32, weights, 1, noArray(), 0, dstMat_fp32, GEMM_2_T);
 
             if (bias)
             {
                 UMat biasOnesMat = UMat::ones(outerSize, 1, umat_blobs[0].type());
                 UMat& biases = umat_blobs[1];
-                cv::gemm(biasOnesMat, biases, 1, dstMat, 1, dstMat, 0);
+                cv::gemm(biasOnesMat, biases, 1, dstMat_fp32, 1, dstMat_fp32, 0);
+            }
+            if (use_half)
+            {
+                convertFp16(srcMat_fp32, srcMat);
+                convertFp16(dstMat_fp32, dstMat);
             }
         }
 

modules/dnn/test/test_caffe_importer.cpp

@@ -490,8 +490,7 @@ INSTANTIATE_TEST_CASE_P(Test_Caffe, opencv_face_detector,
 
 TEST_P(Test_Caffe_nets, FasterRCNN_vgg16)
 {
-    if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD) ||
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
-        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
         throw SkipTestException("");
     static Mat ref = (Mat_<float>(3, 7) << 0, 2, 0.949398, 99.2454, 210.141, 601.205, 462.849,
                                            0, 7, 0.997022, 481.841, 92.3218, 722.685, 175.953,
@@ -502,8 +501,7 @@ TEST_P(Test_Caffe_nets, FasterRCNN_vgg16)
 TEST_P(Test_Caffe_nets, FasterRCNN_zf)
 {
     if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
-        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
         throw SkipTestException("");
     static Mat ref = (Mat_<float>(3, 7) << 0, 2, 0.90121, 120.407, 115.83, 570.586, 528.395,
                                            0, 7, 0.988779, 469.849, 75.1756, 718.64, 186.762,
@@ -514,12 +512,13 @@ TEST_P(Test_Caffe_nets, FasterRCNN_zf)
 TEST_P(Test_Caffe_nets, RFCN)
 {
     if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
-        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
-        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
         throw SkipTestException("");
+    double scoreDiff = (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ? 4e-3 : default_l1;
+    double iouDiff = (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ? 8e-2 : default_lInf;
     static Mat ref = (Mat_<float>(2, 7) << 0, 7, 0.991359, 491.822, 81.1668, 702.573, 178.234,
                                            0, 12, 0.94786, 132.093, 223.903, 338.077, 566.16);
-    testFaster("rfcn_pascal_voc_resnet50.prototxt", "resnet50_rfcn_final.caffemodel", ref);
+    testFaster("rfcn_pascal_voc_resnet50.prototxt", "resnet50_rfcn_final.caffemodel", ref, scoreDiff, iouDiff);
 }
 
 INSTANTIATE_TEST_CASE_P(/**/, Test_Caffe_nets, dnnBackendsAndTargets());