sem_feat, hem_feat and network architecture ?

[soHardToHaveAName]

I've read your paper and the network architecture in the paper. But when I read the code pycenternet_head.py, I find it so hard to understand.

For example,

• what is the meaning of sem_feat? why do you add it to tl_cls_feat, br_cls_feat, tl_pts_feat, br_pts_feat ?
• what is the meaning of hem_feat?
  • why do you extract so many "out"s from it , then merge the "out"s into hem_feat,
  • and again, add it to tl_cls_feat, br_cls_feat, tl_pts_feat, br_pts_feat ?

Seems like your paper lacks explanations for these designs.

Can you provide a detailed vesion of head architecture?

[Duankaiwen]

We refer to the RepPointsV2 for sem_feat and hem_feat, for more details, please refer to their paper: https://arxiv.org/abs/2007.08508