如何生成ground truth？

[liu6381810]

您好
我想请问下如何生成PAF的ground truth
在论文中，如何判断一个点是否在limb上，有一个参数是控制
垂直方向的距离的，但是没有明确说是多少。
我在你的这个函数putVecMaps里看到了一个阈值1
如果我理解不错的话dist = np.absolute(ba_x * bc_y - ba_y * bc_x)计算的是
垂直方向的距离。
那么阈值只有1个像素点不是会很小吗？还是我的理解有问题。
谢谢！！

[liu6381810]

刚才又去看了下代码，发现是除以8以后的1个像素，那对应回原图大概两边加起来就是16？不知道这样理解对不对？

[dragonfly90]

Yes, I followed the original code https://github.com/CMU-Perceptual-Computing-Lab/caffe_train/blob/master/src/caffe/cpm_data_transformer.cpp#L1106. But I don't know if there is a better way. I checked the generated paf in some images. Some are not good, maybe because the human size varies.

for (int g_y = min_y; g_y < max_y; g_y++){
    for (int g_x = min_x; g_x < max_x; g_x++){
      Point2f ba;
      ba.x = g_x - centerA.x;
      ba.y = g_y - centerA.y;
      float dist = std::abs(ba.x*bc.y -ba.y*bc.x);

      // float A = cosine * (g_x - x_p) + sine * (g_y - y_p);
      // float B = sine * (g_x - x_p) - cosine * (g_y - y_p);
      // float judge = A * A / a_sqrt + B * B / b_sqrt;

      if(dist <= thre){
      //if(judge <= 1){
        int cnt = count.at<uchar>(g_y, g_x);
        //LOG(INFO) << "putVecMaps here we start for " << g_x << " " << g_y;
        if (cnt == 0){
          entryX[g_y*grid_x + g_x] = bc.x;
          entryY[g_y*grid_x + g_x] = bc.y;
        }
        else{
          entryX[g_y*grid_x + g_x] = (entryX[g_y*grid_x + g_x]*cnt + bc.x) / (cnt + 1);
          entryY[g_y*grid_x + g_x] = (entryY[g_y*grid_x + g_x]*cnt + bc.y) / (cnt + 1);
          count.at<uchar>(g_y, g_x) = cnt + 1;
        }
      }

    }

[liu6381810]

@dragonfly90 Thanks for you reply.
I generate the ground truth followed the same code.
But the question is for the heat map, the size is 46*46, but only a few points set to values > 0 with all other points set to 0. When I use these data to train the network, it tends to predict all the points to 0 and even so the loss is very low because we use mse loss. So is there any trick in the training procedure?
I would appreciate it if you can give me any advice. Thanks!