search

Joker316701882 / Deep-Image-Matting

This is tensorflow implementation for paper "Deep Image Matting"
626 stars 188 forks source link

fc6 layer & unpooling layers? #2

Closed SarahQiong closed 6 years ago

SarahQiong commented 7 years ago

It looks to me that you are missing the fc6 layer in the encoder and the unpooling layer in the decoder. Please correct me if my understanding is wrong.

p.s. I am interested in implementing deep image matting in tensorflow as well, please send me an email if you are interested in discussing the details of the implementation. Thanks.

Joker316701882 commented 7 years ago

Yes you are right! Seems I missed the fc6, e.g.14th layer. I will correct it. And I replaced unpooling layer by using deconvolution(There is no unpooling ops in tensorflow).It's OK if we have larger training set. This whole decoder part is actually based on my own idea(still in experiment, and it works not bad). The loss architecture is not correct in this code(it may vary based on different understanding). Currently, because of some reason, I have to re-prepare all training data. So the update will be postponed a little. Thank you for your advice!