xwjabc / hed

A PyTorch reimplementation of Holistically-Nested Edge Detection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About eval

BruceYu-Bit opened this issue · comments

想请问在评测代码里面,是否对真值进行了二值化,以及在运行评测标准时,会计算人类的标准,大概score是0.8,这个是如何计算的?还有个问题是bsds500上面的原数据是由七个人标注,然后融合取平均,那在测试集上也是拿平均后的真值去评测的吗?

Hi, thank you for your question. You may refer to this part of the code:

  Z=zeros(size(E)); matchE=Z; matchG=Z; allG=Z;
  for g = 1:n
    [matchE1,matchG1] = correspondPixels(E1,G{g},maxDist);
    matchE = matchE | matchE1>0;
    matchG = matchG + double(matchG1>0);
    allG = allG + G{g};
  end

In this snippet, n=5 (5 ground-truth maps for the same image) and G{g} refers to one ground-truth map. The evaluation procedure will accumulate the matching result for each G{g}. Hope it helps!