Realtime_Multi-Person_Pose_Estimation
Realtime_Multi-Person_Pose_Estimation copied to clipboard
Difference in paper and code
Paper says:
Code is following
https://github.com/CMU-Perceptual-Computing-Lab/caffe_train/blob/master/src/caffe/cpm_data_transformer.cpp
I.e. not max, but min(sum(...),1.0)
Yes, this is actually a bug. It should be min(max(...), 1.0). Thank you for pointing out.
@ZheC thank you.
p.s. Speaking about a bugs - probably you have another one, some PAF could became NaN, I've pointed to the line of code here https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation/issues/102
please check it too.
It should be min(max(...), 1.0)
actually max(...) will be enough since original function can't be more than 1.0
Another difference about PAF.
paper says:
the code is:
If don't update the value of cnt when cnt = 0. The value of cnt always will be 0. And the value of paf will be the last part value, not the average of the affinity fields of all people in the image.
So I think ther is no need to make a special judgment when cnt = 0.
Paper says:
Code is following https://github.com/CMU-Perceptual-Computing-Lab/caffe_train/blob/master/src/caffe/cpm_data_transformer.cpp
I.e. not max, but min(sum(...),1.0)
I can not find which line, the picture you show did not contain max op.