Add DAB-DETR Object detection/segmentation model

[conditionedstimulus]

What does this PR do?

Add DAB-DETR Object detection model. Paper: https://arxiv.org/abs/2201.12329
Original code repo: https://github.com/IDEA-Research/DAB-DETR

Fixes # (issue)
[WIP] This model is part of how DETR models have evolved, alongside DN DETR (not part of this PR), to pave the way for newer and better models like Dino and Stable Dino in object detection

Who can review?

@amyeroberts

[amyeroberts]

Hi @conditionedstimulus, thanks for opening a PR!

Just skimming over the modeling files, it looks like all of the modules are copied from, or can be copied from conditional DETR. Are there any architectural changes this model brings? If not, then all we need to do is convert the checkpoints and upload those to the hub such that they can be loaded in ConditionalDETR directly

[conditionedstimulus]

Hi @conditionedstimulus, thanks for opening a PR!

Just skimming over the modeling files, it looks like all of the modules are copied from, or can be copied from conditional DETR. Are there any architectural changes this model brings? If not, then all we need to do is convert the checkpoints and upload those to the hub such that they can be loaded in ConditionalDETR directly

Hi Amy,

I attached a photo comparing the cross-attention of the decoder in DETR, Conditional DETR, and DAB DETR, as this is the main architectural difference. I copied the code from Conditional DETR because this model is an extension/evolved version of Conditional DETR. I believe it would be cool and useful to include this model in the HF object detection collection.

[amyeroberts]

@conditionedstimulus Thanks for sharing! OK, seems useful to have this available as an option as part of the DETR family in the library. Feel free to ping me when the PR is ready for review.

cc @qubvel for reference