chatop2020
commented
2 years ago ncnn的c++ demo上，我感觉有些问题，我用.net 重写一下，效果还行，你也可以自己改改，主要代码如下：
outImg = extractor.Extract("758");
if (outImg != null)
{
for (int h = 0; h < outImg.H; h++)
{
for (int w = 0; w < outImg.H; w++)
/*
####这里顺便问一下大佬，outImg.H 真的对吗？是不是 你c++代码里笔误了，应该是outImg.W?#####
####我试过outImg.W好像和outImg.H没什么区别######################################
####其实我后来想想，也无所谓，因为进来的图片都是长宽一样的，可能就是因为这个，所以才无所谓吧######
*/
{
//前景概率
int obj_score_index = (0 * outImg.H * outImg.W) + (h * outImg.W) + w;
float obj_score = outImg[obj_score_index];
//解析类别
int category = 0;
float max_score = 0.0f;
for (long i = 0; i < _namesDic.Count; i++)
{
obj_score_index = (int) ((5 + i) * outImg.H * outImg.W) + (h * outImg.W) + w;
float cls_score = outImg[obj_score_index];
if (cls_score > max_score)
{
max_score = cls_score;
category = (int) i;
}
}
float score = (float) (Math.Pow(max_score, 0.4) * Math.Pow(obj_score, 0.6));
// 解析坐标
int x_offset_index = (1 * outImg.H * outImg.W) + (h * outImg.W) + w;
int y_offset_index = (2 * outImg.H * outImg.W) + (h * outImg.W) + w;
int box_width_index = (3 * outImg.H * outImg.W) + (h * outImg.W) + w;
int box_height_index = (4 * outImg.H * outImg.W) + (h * outImg.W) + w;
float x_offset = Tanh(outImg[x_offset_index]);
float y_offset = Tanh(outImg[y_offset_index]);
float box_width = Sigmoid(outImg[box_width_index]);
float box_height = Sigmoid(outImg[box_height_index]);
float cx = (w + x_offset) / outImg.W;
float cy = (h + y_offset) / outImg.H;
int x1 = (int) ((cx - box_width * 0.5) * inputImg.Cols);
int y1 = (int) ((cy - box_height * 0.5) * inputImg.Rows);
int x2 = (int) ((cx + box_width * 0.5) * inputImg.Cols);
int y2 = (int) ((cy + box_height * 0.5) * inputImg.Rows);
var box = new BoundingBox();
box.prob = score;
box.h = (uint) (y2 - y1);
box.w = (uint) (x2 - x1);
box.x = (uint) x1;
box.y = (uint) y1;
box.obj_id = (uint) category;
if (score > threshold && x1 >= 0 && y1 >= 0 && x2 <= inputImg.Cols && y2 <= inputImg.Rows)
{
outBBoxEx.FilterBoundingBoxes.Add(box);
}
else
{
outBBoxEx.OriginalBoundingBoxes.Add(box);
}
}
}
if (outBBoxEx.FilterBoundingBoxes.Count > 0)
{
outBBoxEx.FilterBoundingBoxes.Sort((x, y) => -x.prob.CompareTo(y.prob));
List<picks> picked = new List<picks>();
for (int i = 0; i < outBBoxEx.FilterBoundingBoxes.Count; i++)
{
bool keep = true;
for (int j = 0; j < picked.Count; j++)
{
if (picked[j].Remove) //如果是被过滤掉的，就不执行后续操作
{
continue;
}
var obj1 = outBBoxEx.FilterBoundingBoxes[i];
var obj2 = outBBoxEx.FilterBoundingBoxes[picked[j].Index];
var rect1 = new Rectangle((int) obj1.x, (int) obj1.y, (int) obj1.w, (int) obj1.h);
var rect2 = new Rectangle((int) obj2.x, (int) obj2.y, (int) obj2.w, (int) obj2.h);
var a = bbOverlap(rect1, rect2); //计算两个矩形的交集，如果大于65%的就过滤掉
if (a > 0.65f && obj1.obj_id == obj2.obj_id)
{
keep = false;
break;
}
if (rect2.IntersectsWith(rect1)) //计算两个矩形的并集，如果是包含关系，同时obj1的分数大于obj2,还是同一个类型的，就保留当前的obj1
{
if (obj1.prob > obj2.prob && obj1.obj_id == obj2.obj_id)
{
picked[j].Remove = true;
break;
}
keep = false;
break;
}
}
if (keep)
{
var p = new picks();
p.Index = i;
p.Remove = false;
picked.Add(p);
}
}
for (int i = 0; i < picked.Count; i++)
{
if (picked[i].Remove)
{
continue;
}
var obj = outBBoxEx.FilterBoundingBoxes[picked[i].Index];
if (obj.x >= 0 && obj.y >= 0 && obj.w <= inputImg.Cols && obj.h <= inputImg.Rows)
outBBoxEx.FinishBoundingBoxes.Add(obj);
}
outBBoxEx.FilterBoundingBoxes.Clear();
outBBoxEx.FilterBoundingBoxes = outBBoxEx.FinishBoundingBoxes;
Bitmap img = new Bitmap(filePath);
outBBoxEx.FilePath = filePath;
outBBoxEx.IsImgType = img.RawFormat;
outBBoxEx.Height = inputImg.Rows;
outBBoxEx.Width = inputImg.Cols;
outBBoxEx.PixelFormat = img.PixelFormat;
outBBoxEx.Threshold = threshold;
outBBoxEx.DetectTime = DateTime.Now;
FileInfo fi = new FileInfo(filePath);
outBBoxEx.FileSize = (ulong) fi.Length;
if (createResultImage)
{
darwBox(outBBoxEx, img, filePath + ".jpg");
outBBoxEx.ResultPath = filePath + ".jpg";
}
outBBoxEx.ElapsedMsc = (long) (DateTime.Now - dt1).TotalMilliseconds;
return outBBoxEx;
}
}
clue3clue32
运行ncnn例子，无任何修改，出现这种情况，哪位大佬知道是什么问题