public static KeypointsDetectionInfo[] Decode(Nnet net, int w, int h, double threshold, string[] allowedClasses = null)
{
List <KeypointsDetectionInfo> ret = new List <KeypointsDetectionInfo>();
var inp = net.Nodes.First(z => z.IsInput);
var f1 = net.Nodes.FirstOrDefault(z => z.Dims.Last() == 3);
var snd = net.Nodes.FirstOrDefault(z => z.Dims.Length == 1 && z.ElementType == typeof(float));
if (f1 == null)
{
return(null);
}
var rets1 = net.OutputDatas[f1.Name] as float[];
var scores = net.OutputDatas[snd.Name] as float[];
InternalArray ar = new InternalArray(f1.Dims);
ar.Data = new double[rets1.Length];
for (int i = 0; i < rets1.Length; i++)
{
ar.Data[i] = rets1[i];
}
int cnt = scores.Length;
for (int i = 0; i < cnt; i++)
{
if (scores[i] < 0.9)
{
continue;
}
var kp = new KeypointsDetectionInfo();
ret.Add(kp);
var sub = ar.Get2DImageFrom3DArray(i);
List <Point2f> pp = new List <Point2f>();
for (int j = 0; j < sub.Shape[0]; j++)
{
pp.Add(new Point2f((float)(sub.Get2D(j, 0) / inp.Dims[3] * w), (float)(sub.Get2D(j, 1) / inp.Dims[2] * h)));
}
kp.Points = pp.ToArray();
}
return(ret.ToArray());
}
public static ObjectDetectionInfo[] yoloBoxesDecode(Nnet net, int w, int h, float nms_tresh, double threshold, string[] allowedClasses = null)
{
List <ObjectDetectionInfo> ret = new List <ObjectDetectionInfo>();
var f1 = net.Nodes.FirstOrDefault(z => z.Dims.Last() == 4);
var f2 = net.Nodes.FirstOrDefault(z => z.Dims.Last() > 4);
if (f1 == null || f2 == null)
{
return(null);
}
var nms = Helpers.ReadResource("coco.names").Split(new char[] { '\r', '\n' }, StringSplitOptions.RemoveEmptyEntries).ToArray();
var rets1 = net.OutputDatas[f2.Name] as float[];
var rets3 = net.OutputDatas[f1.Name] as float[];
var dims = net.Nodes.First(z => z.IsInput).Dims;
List <int> indexes = new List <int>();
List <double> confs = new List <double>();
List <int> classes = new List <int>();
Dictionary <int, List <int> > perClassIndexes = new Dictionary <int, List <int> >();
Dictionary <int, List <float> > perClassConfs = new Dictionary <int, List <float> >();
int pos = 0;
var cnt = f2.Dims.Last();
for (int i = 0; i < rets1.Length; i += cnt, pos++)
{
int maxind = -1;
double maxv = double.NaN;
for (int j = 0; j < cnt; j++)
{
if (allowedClasses != null && !allowedClasses.Contains(nms[j]))
{
continue;
}
if (maxind == -1 || rets1[i + j] > maxv)
{
maxv = rets1[i + j];
maxind = j;
}
}
if (maxind != -1 && maxv > threshold)
{
confs.Add(maxv);
classes.Add(maxind);
indexes.Add(pos);
if (!perClassIndexes.ContainsKey(maxind))
{
perClassIndexes.Add(maxind, new List <int>());
perClassConfs.Add(maxind, new List <float>());
}
perClassIndexes[maxind].Add(pos);
perClassConfs[maxind].Add((float)maxv);
}
}
List <int> res = new List <int>();
foreach (var item in perClassIndexes)
{
List <float[]> boxes = new List <float[]>();
for (int i = 0; i < item.Value.Count; i++)
{
var box = rets3.Skip(item.Value[i] * 4).Take(4).ToArray();
boxes.Add(new float[] { box[0], box[1], box[2], box[3], (float)perClassConfs[item.Key][i] });
}
var res2 = Decoders.nms(boxes, nms_tresh);
res.AddRange(res2.Select(z => item.Value[z]));
}
for (int i = 0; i < indexes.Count; i++)
{
if (!res.Contains(indexes[i]))
{
continue;
}
int offset = indexes[i] * 4;
//var box = rets3.Skip(indexes[i] * 4).Take(4).ToArray();
ret.Add(new ObjectDetectionInfo()
{
Class = classes[i],
Conf = (float)confs[i],
Rect = new Rect((int)(rets3[0 + offset] * w),
(int)(rets3[1 + offset] * h),
(int)((rets3[2 + offset] - rets3[0 + offset]) * w),
(int)((rets3[3 + offset] - rets3[1 + offset]) * h)),
Label = nms[classes[i]]
});
}
return(ret.ToArray());
}
public SegmentationDetectionInfo[] Decode(Nnet net, int w, int h, string[] allowedClasses = null)
{
List <SegmentationDetectionInfo> ret = new List <SegmentationDetectionInfo>();
var inp = net.Nodes.First(z => z.IsInput);
var f1 = net.Nodes.FirstOrDefault(z => z.Dims.Last() == inp.Dims.Last());
var snd = net.Nodes.FirstOrDefault(z => z.Dims.Length == 1 && z.ElementType == typeof(float));
var labels = net.Nodes.FirstOrDefault(z => z.Dims.Length == 1 && z.ElementType != typeof(float));
var boxes = net.Nodes.FirstOrDefault(z => z.Dims.Last() == 4);
if (f1 == null)
{
return(null);
}
var nms = Helpers.ReadResource("coco.2.names").Split(new char[] { '\r', '\n' }, StringSplitOptions.RemoveEmptyEntries).ToArray();
var rets1 = net.OutputDatas[f1.Name] as float[];
var labls = net.OutputDatas[labels.Name] as Int64[];
var bxs = net.OutputDatas[boxes.Name] as float[];
var scores = net.OutputDatas[snd.Name] as float[];
InternalArray ar = new InternalArray(f1.Dims);
ar.Data = new double[rets1.Length];
for (int i = 0; i < rets1.Length; i++)
{
ar.Data[i] = rets1[i];
if (rets1[i] > MaskThreshold)
{
ar.Data[i] = 1;
}
else
{
ar.Data[i] = 0;
}
}
int cnt = scores.Length;
for (int i = 0; i < cnt; i++)
{
if (scores[i] < Threshold)
{
continue;
}
var kp = new SegmentationDetectionInfo();
kp.Class = (int)(labls[i]);
kp.Conf = scores[i];
if (labls[i] < nms.Length)
{
kp.Label = nms[labls[i]];
}
else
{
kp.Label = "(unknown)";
}
if (AllowedClasses.Count != 0 && !AllowedClasses.Contains(kp.Label))
{
continue;
}
ret.Add(kp);
double fx = 1f / 900 * w;
double fy = 1f / 600 * h;
Rect rect = new Rect((int)(bxs[i * 4] * fx), (int)(bxs[i * 4 + 1] * fy), (int)((bxs[i * 4 + 2] - bxs[i * 4]) * fx), (int)((bxs[i * 4 + 3] - bxs[i * 4 + 1]) * fy));
kp.Rect = rect;
var sub = ar.Get3DImageFrom4DArray(i);
byte[] arr = new byte[sub.Data.Length];
for (int j = 0; j < sub.Data.Length; j++)
{
arr[j] = (byte)sub.Data[j];
}
var cnt2 = arr.Count(z => z == 1);
if (cnt2 > 0)
{
}
Mat mat = new Mat(600, 900, MatType.CV_8UC1, arr);
kp.Mask = mat;
//mat.SaveImage("test1.jpg");
}
return(ret.ToArray());
}
