search

IDEA-Research / DINO

[ICLR 2023] Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection"
Apache License 2.0
2.15k stars 232 forks source link

Question about the extra norm layer #147

Closed ZRandomize closed 1 year ago

ZRandomize commented 1 year ago

there is an extra norm layer here. Though there is already a layernorm after the FFN, this layer seems improved the final performance. How does it work? I can't find description in the paper.

SlongLiu commented 1 year ago

This implementation inherits from the conditional detr (see here).

We find it helps for the final performance. The extra layernorm can be viewed as an extra linear layer. I suspect the gains come from the extra parameters.

Thanks for providing the interesting question. We will dive into it later.

ZRandomize commented 1 year ago

thx for your answer!