search

seoungwugoh / opn-demo

Onion-Peel Networks for Deep Video Completion, ICCV 2019
https://sites.google.com/view/seoungwugoh
86 stars 20 forks source link

how to train a encoder that encode a input to key and value? Could you detail the loss and other information? #1

Closed biyaa closed 5 years ago

biyaa commented 5 years ago

Hi, I follow your papers "Onion-Peel Networks for Deep Video Completion" and "A Unified Model for Semi-supervised and Interactive Video Object Segmentation using Space-time Memory Networks". The papers all describe the encoder which output key and value, but don't konw how to propel encoder to output key(for addressing) and value (stores detailed information), what is the loss function? how to train encoder? I appreciate it if you reply it . thank you！

tim.credy in China

seoungwugoh commented 5 years ago

Hi, Keys and values are just feature maps, their roles are described according to the network structure. We do not give aditional losses for them, and end-to-end training for losses in the paper are enough to learn them.

biyaa commented 5 years ago

thank you