use OutputArrayOfArrays in net forward interface

It allows umat buffers used in net forward interface
Signed-off-by: Li Peng <peng.li@intel.com>

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

@@ -410,13 +410,13 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
          *  @param outputName name for layer which output is needed to get
          *  @details If @p outputName is empty, runs forward pass for the whole network.
          */
-        CV_WRAP void forward(std::vector<Mat>& outputBlobs, const String& outputName = String());
+        CV_WRAP void forward(OutputArrayOfArrays outputBlobs, const String& outputName = String());
 
         /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
          *  @param outputBlobs contains blobs for first outputs of specified layers.
          *  @param outBlobNames names for layers which outputs are needed to get
          */
-        CV_WRAP void forward(std::vector<Mat>& outputBlobs,
+        CV_WRAP void forward(OutputArrayOfArrays outputBlobs,
                              const std::vector<String>& outBlobNames);
 
         /** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.

modules/dnn/src/dnn.cpp

@@ -1545,7 +1545,7 @@ struct Net::Impl
             CV_Error(Error::StsOutOfRange, "Layer \"" + ld.name + "\" produce only " + toString(ld.outputBlobs.size()) +
                                            " outputs, the #" + toString(pin.oid) + " was requsted");
         }
-        if (preferableBackend != DNN_TARGET_CPU)
+        if (preferableBackend != DNN_BACKEND_DEFAULT)
         {
             // Transfer data to CPU if it's require.
             ld.outputBlobsWrappers[pin.oid]->copyToHost();
@@ -1561,10 +1561,35 @@ struct Net::Impl
         return ld.outputBlobs[pin.oid];
     }
 
+    void getBlob(UMat& umat, const LayerPin& pin)
+    {
+        CV_TRACE_FUNCTION();
+
+        if (!pin.valid())
+            CV_Error(Error::StsObjectNotFound, "Requested blob not found");
+
+        LayerData &ld = layers[pin.lid];
+        if ((size_t)pin.oid >= ld.outputBlobs.size())
+        {
+            CV_Error(Error::StsOutOfRange, "Layer \"" + ld.name + "\" produce only " + toString(ld.outputBlobs.size()) +
+                                           " outputs, the #" + toString(pin.oid) + " was requsted");
+        }
+
+        if (ld.umat_outputBlobs.size() > 0 && !ld.umat_outputBlobs[pin.oid].empty())
+            umat = ld.umat_outputBlobs[pin.oid];
+        else
+            umat = UMat();
+    }
+
     Mat getBlob(String outputName)
     {
         return getBlob(getPinByAlias(outputName));
     }
+
+    void getBlob(UMat& umat, String outputName)
+    {
+        getBlob(umat, getPinByAlias(outputName));
+    }
 };
 
 Net::Net() : impl(new Net::Impl)
@@ -1642,7 +1667,7 @@ Mat Net::forward(const String& outputName)
     return impl->getBlob(layerName);
 }
 
-void Net::forward(std::vector<Mat>& outputBlobs, const String& outputName)
+void Net::forward(OutputArrayOfArrays outputBlobs, const String& outputName)
 {
     CV_TRACE_FUNCTION();
 
@@ -1658,16 +1683,40 @@ void Net::forward(std::vector<Mat>& outputBlobs, const String& outputName)
     LayerPin pin = impl->getPinByAlias(layerName);
     LayerData &ld = impl->layers[pin.lid];
 
-    if (ld.umat_outputBlobs.size() > 0)
+    if (outputBlobs.isUMat())
     {
-        for (int i = 0; i < ld.umat_outputBlobs.size(); i++)
-            ld.umat_outputBlobs[i].copyTo(ld.outputBlobs[i]);
+        if (ld.umat_outputBlobs.size() > 0)
+        {
+            UMat umat;
+            impl->getBlob(umat, layerName);
+            outputBlobs.assign(umat);
+        }
+    }
+    else if (outputBlobs.isMat())
+    {
+        outputBlobs.assign(impl->getBlob(layerName));
+    }
+    else if (outputBlobs.isMatVector())
+    {
+        if (ld.umat_outputBlobs.size() > 0)
+        {
+            for (int i = 0; i < ld.umat_outputBlobs.size(); i++)
+                ld.umat_outputBlobs[i].copyTo(ld.outputBlobs[i]);
+        }
+        std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
+        outputvec = ld.outputBlobs;
+    }
+    else if (outputBlobs.isUMatVector())
+    {
+        if (ld.umat_outputBlobs.size() > 0)
+        {
+            std::vector<UMat> & outputvec = *(std::vector<UMat> *)outputBlobs.getObj();
+            outputvec = ld.umat_outputBlobs;
+        }
     }
-
-    outputBlobs = ld.outputBlobs;
 }
 
-void Net::forward(std::vector<Mat>& outputBlobs,
+void Net::forward(OutputArrayOfArrays outputBlobs,
                   const std::vector<String>& outBlobNames)
 {
     CV_TRACE_FUNCTION();
@@ -1675,7 +1724,7 @@ void Net::forward(std::vector<Mat>& outputBlobs,
     std::vector<LayerPin> pins;
     for (int i = 0; i < outBlobNames.size(); i++)
     {
-       pins.push_back(impl->getPinByAlias(outBlobNames[i]));
+        pins.push_back(impl->getPinByAlias(outBlobNames[i]));
     }
 
     impl->setUpNet(pins);
@@ -1684,11 +1733,14 @@ void Net::forward(std::vector<Mat>& outputBlobs,
 
     impl->forwardToLayer(impl->getLayerData(out.lid));
 
-    outputBlobs.clear();
+    std::vector<Mat> matvec;
     for (int i = 0; i < pins.size(); i++)
     {
-        outputBlobs.push_back(impl->getBlob(pins[i]));
+        matvec.push_back(impl->getBlob(pins[i]));
     }
+
+    std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
+    outputvec = matvec;
 }
 
 void Net::forward(std::vector<std::vector<Mat> >& outputBlobs,