search

google / sg2im

Code for "Image Generation from Scene Graphs", Johnson et al, CVPR 2018
Apache License 2.0
1.29k stars 230 forks source link

Some questions about GraphConvNet #9

Closed shengyuzhang closed 5 years ago

shengyuzhang commented 5 years ago

Hi, I read your paper recently and I have some questions about the "Graph Convolution Network" part, which I am not familiar with:

  1. I noticed this sentence, "Updating object vectors is more complex, since an object may participate in many relationships". I don't quite understand the scope of "many relationships", relationships in a batch? or in the whole dataset? I think your code suggests "in a batch". Is that correct?
  2. I think the graph convolution net is co-trained with the whole model. Is it possible to train graph embeddings as an individual part or as a pre-processed part?

This is actually not a bug but some personal confusions. Actually, I am not a native English speaker, so if you find some expression vague, please feel free to ask me.

Your work is great. :)

shengyuzhang commented 5 years ago

Emmm, I suddenly notice how it works. The scope is one graph grounded with an image.

Sorry to bother. Closed.