94 lines
3.4 KiB
C#
94 lines
3.4 KiB
C#
using System.Reflection;
|
|
using Microsoft.ML.OnnxRuntime;
|
|
using Microsoft.ML.OnnxRuntime.Tensors;
|
|
using OpenCvSharp;
|
|
|
|
namespace CandyboxPlugin.Detectors
|
|
{
|
|
public class CandyDetector
|
|
{
|
|
|
|
private const string MODEL_NAME = "model_transparent_candy_resnet.onnx";
|
|
private readonly InferenceSession _AIsession;
|
|
private readonly string _inputName;
|
|
private readonly int[] _dimensions;
|
|
public CandyDetector()
|
|
{
|
|
var opts = new SessionOptions();
|
|
// get this dll path
|
|
var path = Assembly.GetCallingAssembly().Location;
|
|
var dir = System.IO.Path.GetDirectoryName(path);
|
|
var modelPath = Path.Combine(dir, "models\\" + MODEL_NAME);
|
|
_AIsession = new InferenceSession(modelPath, opts);
|
|
_inputName = _AIsession.InputMetadata.First().Key;
|
|
_dimensions = _AIsession.InputMetadata.First().Value.Dimensions;
|
|
foreach (var input in _AIsession.InputMetadata)
|
|
{
|
|
Console.WriteLine($"Name: {input.Key}");
|
|
Console.WriteLine($" Type: {input.Value.ElementType}");
|
|
Console.WriteLine($" Shape: [{string.Join(", ", input.Value.Dimensions)}]");
|
|
}
|
|
}
|
|
|
|
|
|
public Mat Eval(Mat bmpTest)
|
|
{
|
|
var currentImage = bmpTest;
|
|
|
|
var rightColor = currentImage.CvtColor(ColorConversionCodes.BGR2RGB);
|
|
|
|
// convert to float32
|
|
var floatImage = new Mat();
|
|
rightColor.ConvertTo(floatImage, MatType.CV_32FC3, 1.0 / 255);
|
|
// to array
|
|
|
|
var width = _dimensions[1];
|
|
var height = _dimensions[2];
|
|
|
|
// check if need resize
|
|
if (floatImage.Width != width || floatImage.Height != height)
|
|
Cv2.Resize(floatImage, floatImage, new OpenCvSharp.Size(width, height));
|
|
|
|
|
|
float[] data = new float[3 * width * height];
|
|
int idx = 0;
|
|
var rows = floatImage.Rows;
|
|
var cols = floatImage.Cols;
|
|
for (int y = 0; y < rows; y++)
|
|
{
|
|
for (int x = 0; x < cols; x++)
|
|
{
|
|
Vec3f pixel = floatImage.At<Vec3f>(y, x);
|
|
for (int c = 0; c < 3; c++) // channels
|
|
{
|
|
data[idx++] = pixel[c]; // width*height*channel
|
|
}
|
|
}
|
|
}
|
|
|
|
var inputTensor = new DenseTensor<float>(
|
|
data,
|
|
new int[] { 1, floatImage.Width, floatImage.Height, 3 }
|
|
);
|
|
|
|
var inputs = new List<NamedOnnxValue> {
|
|
NamedOnnxValue.CreateFromTensor("input", inputTensor)
|
|
};
|
|
var sw = System.Diagnostics.Stopwatch.StartNew();
|
|
using var results = _AIsession.Run(inputs);
|
|
sw.Stop();
|
|
Console.WriteLine($"Inference time: {sw.ElapsedMilliseconds} ms");
|
|
var output = results.First().AsEnumerable<float>().ToArray();
|
|
|
|
// convert output to Mat and normalize to 0-255
|
|
var outputMat = new Mat(new OpenCvSharp.Size(width, height), MatType.CV_32FC1);
|
|
outputMat.SetArray<float>(output);
|
|
|
|
outputMat.ConvertTo(outputMat, MatType.CV_8UC1, 255);
|
|
|
|
// resize back to original size
|
|
Cv2.Resize(outputMat, outputMat, new OpenCvSharp.Size(currentImage.Width, currentImage.Height));
|
|
return outputMat;
|
|
}
|
|
}
|
|
} |