search

shihenw / convolutional-pose-machines-release

Code repository for Convolutional Pose Machines, CVPR'16
Other
878 stars 343 forks source link

Meaning of P + 1 part for background #79

Closed sieu-n closed 3 years ago

sieu-n commented 4 years ago

I was reading the paper "Convolutional Pose Machines" and couldn't understand why the model outputted P+1 channels including the channel for the background. What do you intend by predicting the background, and how is the background prediction evaluated?

Sorry if this is an obvious question.

vanheukelomdean commented 4 years ago

I am also confused by this... how is the loss measured for this belief map

alecda573 commented 4 years ago

The background is the P+1 belief map. It is calculated by generating a map as the max across the previous heatmaps and then swapping black for white pixels