//javadoc: NMSBoxesRotated(bboxes, scores, score_threshold, nms_threshold, indices)
public static void NMSBoxesRotated(MatOfRotatedRect bboxes, MatOfFloat scores, float score_threshold, float nms_threshold, MatOfInt indices)
{
if (bboxes != null)
{
bboxes.ThrowIfDisposed();
}
if (scores != null)
{
scores.ThrowIfDisposed();
}
if (indices != null)
{
indices.ThrowIfDisposed();
}
#if ((UNITY_ANDROID || UNITY_IOS || UNITY_WEBGL) && !UNITY_EDITOR) || UNITY_5 || UNITY_5_3_OR_NEWER
Mat bboxes_mat = bboxes;
Mat scores_mat = scores;
Mat indices_mat = indices;
dnn_Dnn_NMSBoxesRotated_12(bboxes_mat.nativeObj, scores_mat.nativeObj, score_threshold, nms_threshold, indices_mat.nativeObj);
return;
#else
return;
#endif
}
// Use this for initialization
void Run()
{
//if true, The error log of the Native side OpenCV will be displayed on the Unity Editor Console.
Utils.setDebugMode(true);
Mat img = Imgcodecs.imread(image_filepath, Imgcodecs.IMREAD_COLOR);
if (img.empty())
{
Debug.LogError(image_filepath + " is not loaded. Please see \"StreamingAssets/dnn/setup_dnn_module.pdf\". ");
img = new Mat(368, 368, CvType.CV_8UC3, new Scalar(0, 0, 0));
}
//Adust Quad.transform.localScale.
gameObject.transform.localScale = new Vector3(img.width(), img.height(), 1);
Debug.Log("Screen.width " + Screen.width + " Screen.height " + Screen.height + " Screen.orientation " + Screen.orientation);
float imageWidth = img.width();
float imageHeight = img.height();
float widthScale = (float)Screen.width / imageWidth;
float heightScale = (float)Screen.height / imageHeight;
if (widthScale < heightScale)
{
Camera.main.orthographicSize = (imageWidth * (float)Screen.height / (float)Screen.width) / 2;
}
else
{
Camera.main.orthographicSize = imageHeight / 2;
}
Net detector = null;
Net recognizer = null;
if (string.IsNullOrEmpty(detectionmodel_filepath) || string.IsNullOrEmpty(recognitionmodel_filepath))
{
Debug.LogError(detectionmodel_filepath + " or " + recognitionmodel_filepath + " is not loaded. Please see \"StreamingAssets/dnn/setup_dnn_module.pdf\". ");
}
else
{
detector = Dnn.readNet(detectionmodel_filepath);
recognizer = Dnn.readNet(recognitionmodel_filepath);
}
if (detector == null || recognizer == null)
{
Imgproc.putText(img, "model file is not loaded.", new Point(5, img.rows() - 30), Imgproc.FONT_HERSHEY_SIMPLEX, 0.7, new Scalar(255, 255, 255), 2, Imgproc.LINE_AA, false);
Imgproc.putText(img, "Please read console message.", new Point(5, img.rows() - 10), Imgproc.FONT_HERSHEY_SIMPLEX, 0.7, new Scalar(255, 255, 255), 2, Imgproc.LINE_AA, false);
}
else
{
TickMeter tickMeter = new TickMeter();
List <Mat> outs = new List <Mat>();
List <string> outNames = new List <string>();
outNames.Add("feature_fusion/Conv_7/Sigmoid");
outNames.Add("feature_fusion/concat_3");
// Create a 4D blob from a frame.
Size inpSize = new Size(inpWidth > 0 ? inpWidth : img.cols(), inpHeight > 0 ? inpHeight : img.rows());
Mat blob = Dnn.blobFromImage(img, 1.0, inpSize, new Scalar(123.68, 116.78, 103.94), true, false); // blobFromImage(frame, blob, 1.0, Size(inpWidth, inpHeight), Scalar(123.68, 116.78, 103.94), true, false);
// Run detection model.
detector.setInput(blob);
tickMeter.start();
detector.forward(outs, outNames);
tickMeter.stop();
Mat scores = outs[0];
Mat geometry = outs[1];
// Decode predicted bounding boxes.
List <RotatedRect> boxes = new List <RotatedRect>();
List <float> confidences = new List <float>();
decodeBoundingBoxes(scores, geometry, confThreshold, boxes, confidences);
// Apply non-maximum suppression procedure.
MatOfRotatedRect boxesMat = new MatOfRotatedRect(boxes.ToArray());
MatOfFloat confidencesMat = new MatOfFloat(confidences.ToArray());
MatOfInt indicesMat = new MatOfInt();
Dnn.NMSBoxesRotated(boxesMat, confidencesMat, confThreshold, nmsThreshold, indicesMat);
List <int> indices = indicesMat.toList();
Point ratio = new Point(img.cols() / inpWidth, img.rows() / inpHeight);
// Render text.
for (int i = 0; i < indices.Count; ++i)
{
RotatedRect box = boxes[indices[i]];
Point[] vertices = new Point[4];
box.points(vertices);
for (int j = 0; j < 4; ++j)
{
vertices[j].x *= ratio.x;
vertices[j].y *= ratio.y;
}
for (int j = 0; j < 4; ++j)
{
Imgproc.line(img, vertices[j], vertices[(j + 1) % 4], new Scalar(0, 255, 0), 1);
}
if (recognizer != null)
{
Mat cropped = new Mat();
fourPointsTransform(img, vertices, cropped);
//Debug.Log(cropped);
Imgproc.cvtColor(cropped, cropped, Imgproc.COLOR_BGR2GRAY);
Mat blobCrop = Dnn.blobFromImage(cropped, 1.0 / 127.5, new Size(), Scalar.all(127.5));
recognizer.setInput(blobCrop);
//Debug.Log(blobCrop);
tickMeter.start();
Mat result = recognizer.forward();
tickMeter.stop();
string wordRecognized;
decodeText(result, out wordRecognized);
Imgproc.putText(img, wordRecognized, vertices[1], Imgproc.FONT_HERSHEY_SIMPLEX, 0.5, new Scalar(255, 0, 0), 1, Imgproc.LINE_AA, false);
Debug.Log(wordRecognized);
cropped.Dispose();
blobCrop.Dispose();
result.Dispose();
}
}
Debug.Log("Inference time, ms: " + tickMeter.getTimeMilli());
for (int i = 0; i < outs.Count; i++)
{
outs[i].Dispose();
}
blob.Dispose();
detector.Dispose();
recognizer.Dispose();
}
Imgproc.cvtColor(img, img, Imgproc.COLOR_BGR2RGB);
Texture2D texture = new Texture2D(img.cols(), img.rows(), TextureFormat.RGBA32, false);
Utils.matToTexture2D(img, texture);
gameObject.GetComponent <Renderer>().material.mainTexture = texture;
Utils.setDebugMode(false);
}
