search

shihenw / convolutional-pose-machines-release

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

kernel size: 9 vs multiple 3s #39

Closed hyqneuron closed 7 years ago

hyqneuron commented 7 years ago

Hi Shih-En,

Thanks for the great work. I've noticed genProto.py generates models with 9x9 kernels whereas the prototxts under model directory contain multiple 3x3 kernels instead. Is it because multiple 3x3 kernels are faster? What are the differences? Thanks!

Yunqing

shihenw commented 7 years ago

Yunqing,

The model with 9x9 kernels were the older version when we trained from scratch. Later on we found training with pre-trained parameters in VGG network in stage 1 of the CPM is beneficial. That's why you see multiple 3x3 there: it's part of VGG net.

hyqneuron commented 7 years ago

That makes sense. Thanks!