Fix padding for average pooling from TensorFlow

diff --git a/modules/dnn/src/layers/pooling_layer.cpp b/modules/dnn/src/layers/pooling_layer.cpp
index 8e5faf9..b451b4b 100644 (file)
--- a/modules/dnn/src/layers/pooling_layer.cpp
+++ b/modules/dnn/src/layers/pooling_layer.cpp
@@ -228,6 +228,7 @@ public:
         const Mat* src, *rois;
         Mat *dst, *mask;
         Size kernel, stride, pad;
+        String padMode;
         int nstripes;
         bool computeMaxIdx;
         std::vector<int> ofsbuf;
@@ -238,7 +239,7 @@ public:
                            computeMaxIdx(0), poolingType(MAX), spatialScale(0) {}
 
         static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask, Size kernel,
-                        Size stride, Size pad, int poolingType, float spatialScale,
+                        Size stride, Size pad, String padMode, int poolingType, float spatialScale,
                         bool computeMaxIdx, int nstripes)
         {
             CV_Assert(src.isContinuous(), dst.isContinuous(),
@@ -257,6 +258,7 @@ public:
             p.kernel = kernel;
             p.stride = stride;
             p.pad = pad;
+            p.padMode = padMode;
             p.nstripes = nstripes;
             p.computeMaxIdx = computeMaxIdx;
             p.poolingType = poolingType;
@@ -336,7 +338,6 @@ public:
                     yend = min(ystart + kernel_h, inp_height + pad_h);
                     srcData = src->ptr<float>(n, c);
                 }
-                int ydelta = yend - ystart;
                 ystart = max(ystart, 0);
                 yend = min(yend, inp_height);
                 float *dstData = dst->ptr<float>(n, c, y0);
@@ -500,15 +501,15 @@ public:
                     }
                 else if (poolingType == AVE)
                 {
+                    bool isSamePad = padMode == "SAME";
                     for( ; x0 < x1; x0++ )
                     {
                         int xstart = x0 * stride_w - pad_w;
                         int xend = min(xstart + kernel_w, inp_width + pad_w);
-                        int xdelta = xend - xstart;
                         xstart = max(xstart, 0);
                         xend = min(xend, inp_width);
-                        float inv_kernel_area = 1.f/(ydelta*xdelta);
-
+                        float inv_kernel_area = isSamePad ? (yend - ystart) * (xend - xstart) : kernel.area();
+                        inv_kernel_area = 1.0 / inv_kernel_area;
 #if CV_SIMD128
                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
                         {
@@ -619,21 +620,21 @@ public:
     {
         const int nstripes = getNumThreads();
         Mat rois;
-        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, padMode, type, spatialScale, computeMaxIdx, nstripes);
     }
 
     void avePooling(Mat &src, Mat &dst)
     {
         const int nstripes = getNumThreads();
         Mat rois, mask;
-        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, padMode, type, spatialScale, computeMaxIdx, nstripes);
     }
 
     void roiPooling(const Mat &src, const Mat &rois, Mat &dst)
     {
         const int nstripes = getNumThreads();
         Mat mask;
-        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
+        PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, padMode, type, spatialScale, computeMaxIdx, nstripes);
     }
 
     virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)

diff --git a/modules/dnn/test/test_tf_importer.cpp b/modules/dnn/test/test_tf_importer.cpp
index 0b4dc64..7ddac16 100644 (file)
--- a/modules/dnn/test/test_tf_importer.cpp
+++ b/modules/dnn/test/test_tf_importer.cpp
@@ -164,6 +164,7 @@ TEST(Test_TensorFlow, pooling)
     runTensorFlowNet("max_pool_even");
     runTensorFlowNet("max_pool_odd_valid");
     runTensorFlowNet("max_pool_odd_same");
+    runTensorFlowNet("ave_pool_same");
 }
 
 TEST(Test_TensorFlow, deconvolution)
@@ -248,6 +249,36 @@ TEST(Test_TensorFlow, MobileNet_SSD)
     normAssert(target[2].reshape(1, 1), output[2].reshape(1, 1), "", 4e-5, 1e-2);
 }
 
+TEST(Test_TensorFlow, Inception_v2_SSD)
+{
+    std::string proto = findDataFile("dnn/ssd_inception_v2_coco_2017_11_17.pbtxt", false);
+    std::string model = findDataFile("dnn/ssd_inception_v2_coco_2017_11_17.pb", false);
+
+    Net net = readNetFromTensorflow(model, proto);
+    Mat img = imread(findDataFile("dnn/street.png", false));
+    Mat blob = blobFromImage(img, 1.0f / 127.5, Size(300, 300), Scalar(127.5, 127.5, 127.5), true, false);
+
+    net.setInput(blob);
+    // Output has shape 1x1xNx7 where N - number of detections.
+    // An every detection is a vector of values [id, classId, confidence, left, top, right, bottom]
+    Mat out = net.forward();
+    out = out.reshape(1, out.total() / 7);
+
+    Mat detections;
+    for (int i = 0; i < out.rows; ++i)
+    {
+        if (out.at<float>(i, 2) > 0.5)
+          detections.push_back(out.row(i).colRange(1, 7));
+    }
+
+    Mat ref = (Mat_<float>(5, 6) << 1, 0.90176028, 0.19872092, 0.36311883, 0.26461923, 0.63498729,
+                                    3, 0.93569964, 0.64865261, 0.45906419, 0.80675775, 0.65708131,
+                                    3, 0.75838411, 0.44668293, 0.45907149, 0.49459291, 0.52197015,
+                                    10, 0.95932811, 0.38349164, 0.32528657, 0.40387636, 0.39165527,
+                                    10, 0.93973452, 0.66561931, 0.37841269, 0.68074018, 0.42907384);
+    normAssert(detections, ref);
+}
+
 OCL_TEST(Test_TensorFlow, MobileNet_SSD)
 {
     throw SkipTestException("TODO: test is failed");