public async Task TestMaskRCNN()
{
using (Tensor imageTensor = ImageIO.ReadTensorFromImageFile <byte>(
"surfers.jpg",
-1,
-1,
0,
1.0f))
using (MaskRcnnInceptionV2Coco model = new MaskRcnnInceptionV2Coco())
{
await model.Init();
MaskRcnnInceptionV2Coco.RecognitionResult[][] results = model.Recognize(imageTensor);
}
}
public void Recognize(Mat m)
{
int[] dim = new int[] { 1, m.Height, m.Width, 3 };
if (_imageTensor == null)
{
_imageTensor = new Tensor(Emgu.TF.DataType.Uint8, dim);
}
else
{
if (!(_imageTensor.Type == Emgu.TF.DataType.Uint8 && Enumerable.SequenceEqual(dim, _imageTensor.Dim)))
{
_imageTensor.Dispose();
_imageTensor = new Tensor(Emgu.TF.DataType.Uint8, dim);
}
}
Emgu.TF.TensorConvert.ReadTensorFromMatBgr(m, _imageTensor);
MaskRcnnInceptionV2Coco.RecognitionResult[] results;
if (_coldSession)
{
//First run of the recognition graph, here we will compile the graph and initialize the session
//This is expected to take much longer time than consecutive runs.
results = _inceptionGraph.Recognize(_imageTensor)[0];
_coldSession = false;
}
//Here we are trying to time the execution of the graph after it is loaded
Stopwatch sw = Stopwatch.StartNew();
results = _inceptionGraph.Recognize(_imageTensor)[0];
sw.Stop();
int goodResultCount = 0;
foreach (var r in results)
{
if (r.Probability > 0.5)
{
float x1 = r.Region[0] * m.Height;
float y1 = r.Region[1] * m.Width;
float x2 = r.Region[2] * m.Height;
float y2 = r.Region[3] * m.Width;
RectangleF rectf = new RectangleF(y1, x1, y2 - y1, x2 - x1);
Rectangle rect = Rectangle.Round(rectf);
rect.Intersect(new Rectangle(Point.Empty, m.Size)); //only keep the region that is inside the image
if (rect.IsEmpty)
{
continue;
}
//draw the rectangle around the region
CvInvoke.Rectangle(m, rect, new Emgu.CV.Structure.MCvScalar(0, 0, 255), 2);
#region draw the mask
float[,] mask = r.Mask;
GCHandle handle = GCHandle.Alloc(mask, GCHandleType.Pinned);
using (Mat mk = new Mat(new Size(mask.GetLength(1), mask.GetLength(0)), Emgu.CV.CvEnum.DepthType.Cv32F, 1, handle.AddrOfPinnedObject(), mask.GetLength(1) * sizeof(float)))
using (Mat subRegion = new Mat(m, rect))
using (Mat maskLarge = new Mat())
using (Mat maskLargeInv = new Mat())
using (Mat largeColor = new Mat(subRegion.Size, Emgu.CV.CvEnum.DepthType.Cv8U, 3))
{
CvInvoke.Resize(mk, maskLarge, subRegion.Size);
//give the mask at least 30% transparency
using (ScalarArray sa = new ScalarArray(0.7))
CvInvoke.Min(sa, maskLarge, maskLarge);
//Create the inverse mask for the original image
using (ScalarArray sa = new ScalarArray(1.0))
CvInvoke.Subtract(sa, maskLarge, maskLargeInv);
//The mask color
largeColor.SetTo(new Emgu.CV.Structure.MCvScalar(255, 0, 0));
CvInvoke.BlendLinear(largeColor, subRegion, maskLarge, maskLargeInv, subRegion);
}
handle.Free();
#endregion
//draw the label
CvInvoke.PutText(m, r.Label, Point.Round(rect.Location), Emgu.CV.CvEnum.FontFace.HersheyComplex, 1.0, new Emgu.CV.Structure.MCvScalar(0, 255, 0), 1);
goodResultCount++;
}
}
String resStr = String.Format("{0} objects detected in {1} milliseconds.", goodResultCount, sw.ElapsedMilliseconds);
if (_renderMat == null)
{
_renderMat = new Mat();
}
m.CopyTo(_renderMat);
//Bitmap bmp = _renderMat.ToBitmap();
if (InvokeRequired)
{
this.Invoke((MethodInvoker)(() =>
{
messageLabel.Text = resStr;
pictureBox.Image = _renderMat;
}));
}
else
{
messageLabel.Text = resStr;
pictureBox.Image = _renderMat;
}
}
