fundamentalvision / Deformable-DETR

Deformable DETR: Deformable Transformers for End-to-End Object Detection.
Apache License 2.0
3.14k stars 513 forks source link

Reference points #233

Open ReefAlturki opened 3 months ago

ReefAlturki commented 3 months ago

Thank you for your great work.

I would like to ask about the rule of reference points, is attention going to be computed only for these reference points, meaning that the reference points are the only locations in the feature map that are going to be updated by attending to relevant neighboring features. I saw a number of youtube videos and tutorials but unfortunately could not understand the rule of reference points.

If I want to only compute the attention on the locations of reference points (while leaving other features in the feature map unchanged) are the codes presented in this paper going to work?

Best wishes,