Geometry-adaptive kernel

leeyeehoo / CSRNet-pytorch

CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes

642 stars 259 forks source link

Geometry-adaptive kernel #65

Open tempdata73 opened 4 years ago

tempdata73 commented 4 years ago

I couln't help but notice that the sigma defined in the code is different from the one mentioned in the paper. The latter is defined as the mean of the distances times beta=0.3, whilst the former is just the sum of the distances times beta=0.1. Does this make a real difference in performance/evaluation? issue

If I generate the density map by following the paper, I get this image, which is clearly different from the other ones... density

jimeffry commented 4 years ago

I guess it is right, because (d1+d2+d3)/3 0.3 = (d1+d2+d3) 0.1