RF-DETR + Dynamic Query for small object detection

[john09282922]

Search before asking

• I have searched the RF-DETR issues and found no similar feature requests.

Description

Adoption of Dynamic Query. I think RF-DETR follows a DETR-style pipeline (backbone → transformer decoder with queries → heads), so the query generator is a modular insertion point. Possible to graft Dynamic Query (DQ) from DQ-DETR—i.e., a per-image adaptive query count and density-guided query initialization—by replacing the fixed query embedding + fixed reference points with a learned counting head + density head that drive per-image query sampling. Decoder, heads, and matching remain compatible after minor masking changes. I am not sure this idea is good for small object detection? what do you think? and it is not effect on inference speed.

Use case

No response

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[isaacrob-roboflow]

rf-detr is a two stage detr, so its queries are initialized from the backbone. this is a simplification of the dq-detr design you've linked above .. the dq may work better, which would be interesting to study, but it is not the case that it is currently using fixed queries