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
2.27k stars 257 forks source link

a question about the backbone of covnext-xl #132

Open bartbuaa opened 2 years ago

bartbuaa commented 2 years ago

It seems that various backbones obviously affect the final MAP score. Since ConvNeXt-XL surpass Swin-L on ImageNet-22k, it is curious that have you ever tried to utilize it for improving. I notice that your code includes ConvNeXt-XL indeed. Could you offer some details about the MAP score on COCO with DINO ConvNeXt-XL.

It is so appreciated!

FengLi-ust commented 2 years ago

In our preliminary experiments, SwinL outperforms ConvNex. But we did not tune the hyperparameters, like the weight decay. Therefore, we are not sure about the performance of ConvNext.