Questions about location tokens #4

Activity

1. kahnchana commented on Jul 17, 2023

   

   kahnchana

   on Jul 17, 2023

   More actions

   Same here, I am trying to figure out what they are.

   Are they a fixed grid or some learnable parameter?
2. kahnchana commented on Aug 15, 2023

   

   kahnchana

   on Aug 15, 2023

   More actions

   I think it appears to be the same grid like structure used in deformable DETR. Basically it's a uniform grid across image coordinates, and each grid centre is used as an anchor, over which the model regresses the deviation of correct bbox.
3. SxJyJay commented on Jan 19, 2024

   

   SxJyJay

   on Jan 19, 2024

   More actions

   Basically it's a uniform grid across image coordinates, and each grid centre is used as an anchor, over which the model regresses the deviation of correct bbox.

   Hi, I have a similar problem with you. If VisionLLM uses the Deformable DETR-like decoder, and object queries act as positional anchors, the Hungarian matching is required to assign GT boxes to object queries. However, the authors don't mention that in the paper. What do you think of the possible training details of these object queries?

   👍React with 👍1Reacted by Keumgang Cha
4. chagmgang commented on Feb 6, 2024

   

   chagmgang

   on Feb 6, 2024

   Last edited by chagmgang

   More actions

   Basically it's a uniform grid across image coordinates, and each grid centre is used as an anchor, over which the model regresses the deviation of correct bbox.

   Hi, I have a similar problem with you. If VisionLLM uses the Deformable DETR-like decoder, and object queries act as positional anchors, the Hungarian matching is required to assign GT boxes to object queries. However, the authors don't mention that in the paper. What do you think of the possible training details of these object queries?

   Also, Same here.

   #7 (comment)

   maybe below link will be helpful

   https://openreview.net/forum?id=Vx1JadlOIt&noteId=616Bhd6O5S

   maybe below image will be helpful too

   

   👍React with 👍1Reacted by YangJiao