public static Tensor ProcessPrediction(int inputSize, YOLOv4.Output modelOutput, int classCount, ReadOnlySpan <int> strides, Tensor <int> anchors, ReadOnlySpan <float> xyScale, float scoreThreshold)
{
var bboxTensors = new List <Tensor>();
var probTensors = new List <Tensor>();
foreach (var(scaleIndex, featureMap) in Tools.Enumerate(modelOutput.SSBox, modelOutput.MBBox, modelOutput.LBBox))
{
var outputTensors = Decode(featureMap,
outputSize: inputSize / strides[scaleIndex],
classCount: classCount,
strides: strides,
anchors: anchors,
scaleIndex: scaleIndex,
xyScale: xyScale);
bboxTensors.Add(outputTensors.xywh);
probTensors.Add(outputTensors.prob);
}
var bbox = tf.concat(bboxTensors.ToArray(), axis: 1);
var prob = tf.concat(probTensors.ToArray(), axis: 1);
var(boxes, conf) = FilterBoxes(bbox, prob,
scoreThreshold: scoreThreshold,
inputShape: tf.constant(new[] { inputSize, inputSize }));
return(tf.concat(new[] { boxes, conf }, axis: -1));
}
SelectBoxes(YOLOv4.Output featureMaps, int inputSize, int classCount,
ReadOnlySpan <int> strides, Tensor <int> anchors,
ReadOnlySpan <float> xyScale,
float scoreThreshold = 0.2f)
{
var pred = ProcessPrediction(inputSize: inputSize, featureMaps,
classCount: classCount,
strides: strides,
anchors: anchors,
xyScale: xyScale,
scoreThreshold: scoreThreshold);
var boxes = pred[.., .., 0..4];
public static Model CreateSaveable(int inputSize, Tensor input, YOLOv4.Output featureMaps,
int classCount,
ReadOnlySpan <int> strides, Tensor <int> anchors,
ReadOnlySpan <float> xyScale, float scoreThreshold)
{
var suppression = SelectBoxes(featureMaps, inputSize: inputSize, classCount: classCount,
strides: strides, anchors: anchors,
xyScale: xyScale,
scoreThreshold: scoreThreshold);
return(new Model(new { inputs = input, outputs = new PythonList <Tensor> {
suppression.Boxes, suppression.Scores, suppression.Classes, suppression.Detections,
} }.AsKwArgs()));
}
public static ObjectDetectionResult[] DetectRaw(Model rawDetector,
Size supportedSize, int classCount,
Image <Rgb24> image,
ReadOnlySpan <int> strides, Tensor <int> anchors,
ReadOnlySpan <float> xyScale,
float scoreThreshold = 0.2f)
{
if (rawDetector is null)
{
throw new ArgumentNullException(nameof(rawDetector));
}
if (image is null)
{
throw new ArgumentNullException(nameof(image));
}
var input = ImageTools.YoloPreprocess(new ObjectDetectionDataset.ClrEntry {
Image = image.Clone(),
}, supportedSize);
var images = input.Image[np.newaxis, np.rest_of_the_axes].AsArray();
IList <Tensor> prediction = rawDetector.__call__(images);
Debug.Assert(prediction.Count == 3);
var output = new YOLOv4.Output {
SSBox = prediction[0],
MBBox = prediction[1],
LBBox = prediction[2],
};
var suppression = SelectBoxes(output, inputSize: supportedSize.Width, classCount: classCount,
strides: strides, anchors: anchors,
xyScale: xyScale,
scoreThreshold: scoreThreshold);
ndarray <float> boxs = suppression.Boxes.numpy();
ndarray <float> scores = suppression.Scores.numpy();
ndarray <long> classes = suppression.Classes.numpy();
ndarray <int> detections = suppression.Detections.numpy();
return(ObjectDetectionResult.FromCombinedNonMaxSuppressionBatch(
boxs, scores, classes, detections[0].AsScalar()));
}
