add scale factor to DB demo.

modules/dnn/src/dnn_common.hpp

@@ -154,6 +154,21 @@ static inline std::string toString(const Mat& blob, const std::string& name = st
     return ss.str();
 }
 
+// Scalefactor is a common parameter used for data scaling. In OpenCV, we often use Scalar to represent it.
+// Because 0 is meaningless in scalefactor.
+// If the scalefactor is (x, 0, 0, 0), we convert it to (x, x, x, x). The following func will do this hack.
+static inline Scalar_<double> broadcastRealScalar(const Scalar_<double>& _scale)
+{
+    Scalar_<double> scale = _scale;
+    if (scale[1] == 0 && scale[2] == 0 && scale[3] == 0)
+    {
+        CV_Assert(scale[0] != 0 && "Scalefactor of 0 is meaningless.");
+        scale = Scalar_<double>::all(scale[0]);
+    }
+
+    return scale;
+}
+
 
 CV__DNN_INLINE_NS_END
 

modules/dnn/src/model.cpp

@@ -21,7 +21,7 @@ struct Model::Impl
 
     Size   size;
     Scalar mean;
-    double  scale = 1.0;
+    Scalar scale = Scalar::all(1.0);
     bool   swapRB = false;
     bool   crop = false;
     Mat    blob;
@@ -60,7 +60,7 @@ public:
     {
         size = size_;
         mean = mean_;
-        scale = scale_;
+        scale = Scalar::all(scale_);
         crop = crop_;
         swapRB = swapRB_;
     }
@@ -75,7 +75,7 @@ public:
         mean = mean_;
     }
     /*virtual*/
-    void setInputScale(double scale_)
+    void setInputScale(const Scalar& scale_)
     {
         scale = scale_;
     }
@@ -97,7 +97,17 @@ public:
         if (size.empty())
             CV_Error(Error::StsBadSize, "Input size not specified");
 
-        blob = blobFromImage(frame, scale, size, mean, swapRB, crop);
+        Image2BlobParams param;
+        param.scalefactor = scale;
+        param.size = size;
+        param.mean = mean;
+        param.swapRB = swapRB;
+        if (crop)
+        {
+            param.paddingmode = DNN_PMODE_CROP_CENTER;
+        }
+        Mat blob = dnn::blobFromImageWithParams(frame, param); // [1, 10, 10, 4]
+
         net.setInput(blob);
 
         // Faster-RCNN or R-FCN
@@ -162,9 +172,11 @@ Model& Model::setInputMean(const Scalar& mean)
     return *this;
 }
 
-Model& Model::setInputScale(double scale)
+Model& Model::setInputScale(const Scalar& scale_)
 {
     CV_DbgAssert(impl);
+
+    Scalar scale = broadcastRealScalar(scale_);
     impl->setInputScale(scale);
     return *this;
 }
@@ -1358,7 +1370,7 @@ struct TextDetectionModel_DB_Impl : public TextDetectionModel_Impl
     {
         CV_TRACE_FUNCTION();
         std::vector< std::vector<Point2f> > results;
-
+        confidences.clear();
         std::vector<Mat> outs;
         processFrame(frame, outs);
         CV_Assert(outs.size() == 1);
@@ -1385,7 +1397,8 @@ struct TextDetectionModel_DB_Impl : public TextDetectionModel_Impl
             std::vector<Point>& contour = contours[i];
 
             // Calculate text contour score
-            if (contourScore(binary, contour) < polygonThreshold)
+            float score = contourScore(binary, contour);
+            if (score < polygonThreshold)
                 continue;
 
             // Rescale
@@ -1398,6 +1411,11 @@ struct TextDetectionModel_DB_Impl : public TextDetectionModel_Impl
 
             // Unclip
             RotatedRect box = minAreaRect(contourScaled);
+            float minLen = std::min(box.size.height/scaleWidth, box.size.width/scaleHeight);
+
+            // Filter very small boxes
+            if (minLen < 3)
+                continue;
 
             // minArea() rect is not normalized, it may return rectangles with angle=-90 or height < width
             const float angle_threshold = 60;  // do not expect vertical text, TODO detection algo property
@@ -1422,10 +1440,12 @@ struct TextDetectionModel_DB_Impl : public TextDetectionModel_Impl
                 approx.emplace_back(vertex[j]);
             std::vector<Point2f> polygon;
             unclip(approx, polygon, unclipRatio);
+            if (polygon.empty())
+                continue;
             results.push_back(polygon);
+            confidences.push_back(score);
         }
 
-        confidences = std::vector<float>(contours.size(), 1.0f);
         return results;
     }
 
@@ -1458,7 +1478,10 @@ struct TextDetectionModel_DB_Impl : public TextDetectionModel_Impl
     {
         double area = contourArea(inPoly);
         double length = arcLength(inPoly, true);
-        CV_Assert(length > FLT_EPSILON);
+
+        if(length == 0.)
+            return;
+
         double distance = area * unclipRatio / length;
 
         size_t numPoints = inPoly.size();