Fix multiple inputs models from Caffe.

Fixed Concat optimization.

modules/dnn/src/caffe/caffe_importer.cpp

@@ -318,8 +318,12 @@ public:
 
             if (type == "Input")
             {
-                addedBlobs.push_back(BlobNote(name, 0, netInputs.size()));
-                netInputs.push_back(name);
+                for (int outNum = 0; outNum < layer.top_size(); outNum++)
+                {
+                    addOutput(layer, 0, outNum);
+                    addedBlobs.back().outNum = netInputs.size();
+                    netInputs.push_back(addedBlobs.back().name);
+                }
                 continue;
             }
 

modules/dnn/src/dnn.cpp

@@ -275,6 +275,16 @@ struct DataLayer : public Layer
         outNames.assign(names.begin(), names.end());
     }
 
+    bool getMemoryShapes(const std::vector<MatShape> &inputs,
+                         const int requiredOutputs,
+                         std::vector<MatShape> &outputs,
+                         std::vector<MatShape> &internals) const
+    {
+        CV_Assert(inputs.size() == requiredOutputs);
+        outputs.assign(inputs.begin(), inputs.end());
+        return false;
+    }
+
 private:
     std::vector<String> outNames;
 };
@@ -1184,7 +1194,7 @@ struct Net::Impl
                                layers[ld.inputBlobsId[i].lid].getLayerInstance()->name.c_str(),
                                inp_i_data->getLayerInstance()->name.c_str()));
 
-                        if(inp_i_data->skipFlags[DNN_BACKEND_DEFAULT])
+                        if(inp_i_data->skipFlags[DNN_BACKEND_DEFAULT] || inp_i_data->consumers.size() != 1)
                             break;
                         realinputs[i] = pin;
                     }
@@ -1206,6 +1216,14 @@ struct Net::Impl
                             Mat& curr_output = inp_i_data->outputBlobs[pin.oid];
                             CV_Assert(output_slice.isContinuous() && output_slice.size == curr_output.size);
                             curr_output = output_slice;
+
+                            pin = ld.inputBlobsId[i];
+                            inp_i_data = &layers[pin.lid];
+                            for (int j = 0; j < inp_i_data->consumers.size(); ++j)
+                            {
+                                LayerPin consumer = inp_i_data->consumers[j];
+                                layers[consumer.lid].inputBlobs[consumer.oid] = &curr_output;
+                            }
                         }
                         ld.skipFlags[DNN_BACKEND_DEFAULT] = true;
                         printf_(("\toptimized out Concat layer %s\n", concatLayer->name.c_str()));
@@ -1235,7 +1253,9 @@ struct Net::Impl
 
         blobManager.reset();
         backendWrappers.clear();
-        blobManager.addReference(LayerPin(0, 0));
+        // Fake references to input blobs.
+        for (int i = 0; i < layers[0].outputBlobs.size(); ++i)
+            blobManager.addReference(LayerPin(0, i));
         for (it = layers.begin(); it != layers.end(); ++it)
         {
             const LayerData& ld = it->second;

modules/dnn/test/test_caffe_importer.cpp

@@ -280,4 +280,31 @@ TEST(Reproducibility_DenseNet_121, Accuracy)
     normAssert(out, ref);
 }
 
+TEST(Test_Caffe, multiple_inputs)
+{
+    const string proto = findDataFile("dnn/layers/net_input.prototxt", false);
+    Net net = readNetFromCaffe(proto);
+
+    Mat first_image(10, 11, CV_32FC3);
+    Mat second_image(10, 11, CV_32FC3);
+    randu(first_image, -1, 1);
+    randu(second_image, -1, 1);
+
+    first_image = blobFromImage(first_image);
+    second_image = blobFromImage(second_image);
+
+    Mat first_image_blue_green = slice(first_image, Range::all(), Range(0, 2), Range::all(), Range::all());
+    Mat first_image_red = slice(first_image, Range::all(), Range(2, 3), Range::all(), Range::all());
+    Mat second_image_blue_green = slice(second_image, Range::all(), Range(0, 2), Range::all(), Range::all());
+    Mat second_image_red = slice(second_image, Range::all(), Range(2, 3), Range::all(), Range::all());
+
+    net.setInput(first_image_blue_green, "old_style_input_blue_green");
+    net.setInput(first_image_red, "different_name_for_red");
+    net.setInput(second_image_blue_green, "input_layer_blue_green");
+    net.setInput(second_image_red, "old_style_input_red");
+    Mat out = net.forward();
+
+    normAssert(out, first_image + second_image);
+}
+
 }

modules/dnn/test/test_layers.cpp

@@ -274,6 +274,48 @@ OCL_TEST(Layer_Test_Concat, Accuracy)
     testLayerUsingCaffeModels("layer_concat", DNN_TARGET_OPENCL);
 }
 
+TEST(Layer_Test_Fused_Concat, Accuracy)
+{
+    // Test case
+    // input
+    //   |
+    //   v
+    // some_layer
+    // |   |
+    // v   v
+    // concat
+    Net net;
+    int interLayer;
+    {
+        LayerParams lp;
+        lp.type = "AbsVal";
+        lp.name = "someLayer";
+        interLayer = net.addLayerToPrev(lp.name, lp.type, lp);
+    }
+    {
+        LayerParams lp;
+        lp.set("axis", 1);
+        lp.type = "Concat";
+        lp.name = "testConcat";
+        int id = net.addLayer(lp.name, lp.type, lp);
+        net.connect(interLayer, 0, id, 0);
+        net.connect(interLayer, 0, id, 1);
+    }
+    int shape[] = {1, 2, 3, 4};
+    Mat input(4, shape, CV_32F);
+    randu(input, 0.0f, 1.0f);  // [0, 1] to make AbsVal an identity transformation.
+
+    net.setInput(input);
+    Mat out = net.forward();
+
+    normAssert(slice(out, Range::all(), Range(0, 2), Range::all(), Range::all()), input);
+    normAssert(slice(out, Range::all(), Range(2, 4), Range::all(), Range::all()), input);
+
+    //
+
+    testLayerUsingCaffeModels("layer_concat_optim", DNN_TARGET_CPU, true, false);
+}
+
 TEST(Layer_Test_Eltwise, Accuracy)
 {
     testLayerUsingCaffeModels("layer_eltwise");