How to use the appearance features of subject and object?

[fangfchen]

In the paper, the appearance features of subject and object (not union) will also be fed to the DR-Net. But in the code, I find it not be used. Only the one-hot vectors of subject and object are used. Dose the task 1 not need to use it? Since only the model of task 1 is open, how are the appearance features of subject and object used in task 2 and 3? Using VGG and FC+softmax to get a 100-dim vector?

[sunbear616]

hello, I get a same question, do you solve it? I am trying to make a detection network with this appearance features. but I didn't find in this code