public IList <BoundingBox> ParseOutputs(float[] yoloModelOutputs, float threshold = .2F)
{
var boxes = new List <BoundingBox>();
for (int cy = 0; cy < RowCount; cy++)
{
for (int cx = 0; cx < ColCount; cx++)
{
for (int b = 0; b < BoxesPerCell; b++)
{
var channel = (b * (ClassCount + BoxInfoCount));
var tx = yoloModelOutputs[GetOffset(cx, cy, channel)];
var ty = yoloModelOutputs[GetOffset(cx, cy, channel + 1)];
var tw = yoloModelOutputs[GetOffset(cx, cy, channel + 2)];
var th = yoloModelOutputs[GetOffset(cx, cy, channel + 3)];
var tc = yoloModelOutputs[GetOffset(cx, cy, channel + 4)];
var x = (cx + MathMethods.Sigmoid(tx)) * CellWidth;
var y = (cy + MathMethods.Sigmoid(ty)) * CellHeight;
var width = (float)Math.Exp(tw) * CellWidth * anchors[b * 2];
var height = (float)Math.Exp(th) * CellHeight * anchors[b * 2 + 1];
var confidence = MathMethods.Sigmoid(tc);
if (confidence < threshold)
{
continue;
}
var classes = new float[ClassCount];
var classOffset = channel + BoxInfoCount;
for (int i = 0; i < ClassCount; i++)
{
classes[i] = yoloModelOutputs[GetOffset(cx, cy, i + classOffset)];
}
var results = MathMethods.Softmax(classes)
.Select((v, iexp) => new { Value = v, Index = iexp });
var labelClass = results.OrderByDescending(r => r.Value).First().Index;
var scoreClass = results.OrderByDescending(r => r.Value).First().Value *confidence;
var testSum = results.Sum(r => r.Value);
if (scoreClass > threshold)
{
boxes.Add(new BoundingBox()
{
Confidence = scoreClass,
X = (x - width / 2),
Y = (y - height / 2),
Width = width,
Height = height,
Label = Enum.GetName(typeof(ObjectLabels), labelClass).ToLower()
});
}
}
}
}
var filteredBoxes = NonMaxSuppression(boxes, 5, .5F);
return(filteredBoxes);
}