public Tuple <Rect[], float[], int[]> boxesDecode(Mat mat1)
{
var f1 = _nodes.FirstOrDefault(z => z.Tags.Contains("conf"));
var f2 = _nodes.FirstOrDefault(z => z.Tags.Contains("loc"));
if (f1 == null || f2 == null)
{
return(null);
}
var rets1 = OutputDatas[f2.Name] as float[];
var rets3 = OutputDatas[f1.Name] as float[];
var dims = _nodes.First(z => z.IsInput).Dims;
var sz = new System.Drawing.Size(dims[3], dims[2]);
if (dims[2] == -1)
{
sz.Height = mat1.Height;
sz.Width = mat1.Width;
}
string key = $"{sz.Width}x{sz.Height}";
if (!Decoders.allPriorBoxes.ContainsKey(key))
{
var pd = Decoders.PriorBoxes2(sz.Width, sz.Height);;
Decoders.allPriorBoxes.Add(key, pd);
}
var prior_data = Decoders.allPriorBoxes[key];
return(Decoders.boxesDecode(mat1.Size(), rets3, rets1, sz, prior_data, visTresh));
}
public static Tuple <Rect[], float[], int[]> boxesDecode(OpenCvSharp.Size matSize, float[] confd, float[] locd, System.Drawing.Size sz, float[][] prior_data, float vis_thresh = 0.5f)
{
Stopwatch sw = Stopwatch.StartNew();
if (confd == null || locd == null)
{
return(null);
}
List <float[]> loc = new List <float[]>();
List <float> scores = new List <float>();
List <int> winners = new List <int>();
float[] variances = new float[] { 0.1f, 0.2f };
var nnInputWidth = sz.Width;
var nnInputHeight = sz.Height;
float wz1 = nnInputWidth;
float hz1 = nnInputHeight;
float[] scale = new float[] { (float)nnInputWidth, (float)nnInputHeight, (float)nnInputWidth, (float)nnInputHeight };
float koef = wz1 / (float)(matSize.Width);
float koef2 = hz1 / (float)(matSize.Height);
float[] resize = new float[] { koef, koef2 };
var rets3 = confd;
var rets1 = locd;
for (var i = 0; i < rets1.Length; i += 4)
{
loc.Add(new float[] { rets1[i + 0], rets1[i + 1], rets1[i + 2], rets1[i + 3] });
}
int numClasses = rets3.Length / (rets1.Length / 4);
for (var i = 0; i < rets3.Length; i += numClasses)
{
if (numClasses > 2)
{
//first class - background usually
float maxj = rets3[i + 1];
int mind = 1;
for (int j = 2; j < numClasses; j++)
{
if (rets3[i + j] > maxj)
{
maxj = rets3[i + j];
mind = j;
}
}
//var sub = rets3.Skip(i + 1).Take(numClasses - 1).Select((v, ii) => new Tuple<int, float>(ii, v)).OrderByDescending(z => z.Item2).First();
winners.Add(mind - 1);
scores.Add(maxj);
}
else
{
scores.Add(rets3[i + 1]);
}
}
var boxes = Decoders.decode(loc, prior_data, variances);
for (var i = 0; i < boxes.Count(); i++)
{
boxes[i][0] = (boxes[i][0] * scale[0]) / resize[0];
boxes[i][1] = (boxes[i][1] * scale[1]) / resize[1];
boxes[i][2] = (boxes[i][2] * scale[2]) / resize[0];
boxes[i][3] = (boxes[i][3] * scale[3]) / resize[1];
}
float[] scale1 = new float[] { wz1, hz1, wz1, hz1, wz1, hz1, wz1, hz1, wz1, hz1 };
float confidence_threshold = 0.2f;
List <int> inds = new List <int>();
for (var i = 0; i < scores.Count(); i++)
{
if (scores[i] > confidence_threshold)
{
inds.Add(i);
}
}
List <float[]> boxes2 = new List <float[]>();
for (var i = 0; i < inds.Count(); i++)
{
boxes2.Add(boxes[inds[i]]);
}
boxes = boxes2.ToArray();
List <float> scores2 = new List <float>();
for (var i = 0; i < inds.Count(); i++)
{
scores2.Add(scores[inds[i]]);
}
scores = scores2;
List <int> winners2 = new List <int>();
if (numClasses > 2)
{
for (var i = 0; i < inds.Count(); i++)
{
winners2.Add(winners[inds[i]]);
}
winners = winners2;
}
var order = Decoders.sort_indexes(scores);
List <float[]> boxes3 = new List <float[]>();
for (var i = 0; i < order.Count(); i++)
{
boxes3.Add(boxes[order[i]]);
}
boxes = boxes3.ToArray();
List <float> scores3 = new List <float>();
for (var i = 0u; i < order.Count(); i++)
{
scores3.Add(scores[order[i]]);
}
scores = scores3;
if (numClasses > 2)
{
List <int> winners3 = new List <int>();
for (var i = 0; i < inds.Count(); i++)
{
winners3.Add(winners[order[i]]);
}
winners = winners3;
}
//2. nms
List <float[]> dets = new List <float[]>();
for (var i = 0; i < boxes.Count(); i++)
{
if (numClasses > 2)
{
dets.Add(new float[] { boxes[i][0], boxes[i][1], boxes[i][2], boxes[i][3], scores[i], winners[i] });
}
else
{
dets.Add(new float[] { boxes[i][0], boxes[i][1], boxes[i][2], boxes[i][3], scores[i] });
}
}
var keep = Decoders.nms(dets, 0.4f);
List <float[]> dets2 = new List <float[]>();
for (var i = 0u; i < keep.Count(); i++)
{
dets2.Add(dets[keep[i]]);
}
dets = dets2;
List <Rect> detections = new List <Rect>();
//float vis_thresh = 0.2f;
List <int> indexMap = new List <int>();
//List<float[]> odets = new List<float[]>();
List <float> oscores = new List <float>();
List <int> owin = new List <int>();
for (var i = 0; i < dets.Count(); i++)
{
var aa = dets[i];
if (aa[4] < vis_thresh)
{
continue;
}
detections.Add(new Rect((int)(aa[0]), (int)(aa[1]), (int)(aa[2] - aa[0]), (int)(aa[3] - aa[1])));
indexMap.Add(i);
//oscores.Add(scores3[i]);
oscores.Add(aa[4]);
if (numClasses > 2)
{
owin.Add((int)aa[5]);
}
}
/* for (var i = 0; i < dets.Count(); i++)
* {
* odets.Add(dets[i]);
* }*/
sw.Stop();
var ret = new Tuple <Rect[], float[], int[]>(detections.ToArray(), oscores.ToArray(), numClasses > 2 ? owin.ToArray() : null);
return(ret);
}
