search

michalfaber / keras_Realtime_Multi-Person_Pose_Estimation

Keras version of Realtime Multi-Person Pose Estimation project
Other
780 stars 372 forks source link

about PAF #82

Open juzhitao opened 6 years ago

juzhitao commented 6 years ago

I am green hand in AI, I read this paper and your code, I am confuse about PAF, from code output_blobs[0] is paf, the shape is WH2*keypoint, what is each value mean? it is calculate by Eq. 10 in paper? and what is d1 and d2 means in Eq. 10?

michalfaber commented 6 years ago

Hi @juzhitao The shape of PAF output is 8 times smaller than the scaled input image because of downsampling by 3 max pool layers in model architecture so the image of size 368x368 (config param: boxsize) gives paf and heatmaps of size 46x46. As for the Eq 10 in paper, we use only a simple approximation which is sampling PAF map along the line between candidate body parts d1 and d2. (d1 and d2 are locations x,y)