Feng Li
Have you modified something in the code or in the running command?
The idea is similar. You can add noise to the 2D pose coordinates and train the model to regress the ground truth.
You can refer to our [extended version](https://arxiv.org/abs/2203.01305) that describes how to add DN to vanillar DETR.
Yes, and we have posted the ResNet-50 training curve in our paper.
Hey, thanks for your interest in our work. We are glad to see that you implement DINO based on DN-DETR. Is there any difference between your re-implemented DINO and our...
That's great. Looks like you are comparing with the old DINO in your table. You can initialize the parameters as the new DINO, so you can achieve around 49.0 in...
I just have a quick look. The lr_drop should be set to 11.
@Vallum Hey, you can pull request so I can merge your code. We can also have discussions if you meet problems.
Yes, we have tried in our extended DN-DETR version. Please refer to our [new arxiv paper](https://arxiv.org/pdf/2203.01305.pdf)
您好，我们的实现是可以支持batchsize>1的情况的。您可以设置batchsize=2在这里debug一下应该就明白原理了。