Hao Zhang
But if you only use 4 negative samples you can get a hit rate of 100% in top5 ranking.
Please refer to #43 .
@dannyhung1128 Thank you for your interesting question. We are sorry that we have not reported the gain of CDN without query selection till now. We will consider adding this result...
You can refer to this [line](https://github.com/IDEACVR/DINO/blob/8758cf02146f306dc36babab4fff1f09c114c682/models/dino/dn_components.py#L43) which obtain the number of dn groups by considering the number of objects in images.
@fushh Your proposed three ideas are interesting. The first and second looks like the original dn-detr and dynamic dn respectively. Dynamic dn can help to add more dn queries, so...
Say you have an instance mask represented as a polygon [(x_0, y_0), (x_1, y_1), ..., (x_N, y_N)]. You can have the upper left corner of the bbox as (min{x_i}, min{y_i})...
Yes, the extra one dimension does represent the "no object".
In my understanding, sigmoid is more suitable for multi-class classification. When the model is not sure which of two classes an object belongs to, it can predict both so that...
We have tried some strong augmentations such as adjusting contrast and brightness for our Swin-L model pre-trained on o365, but we do not see much gain. You should insert your...
We adopted the same learning rate schedule as previous DETR-like works to suddenly drop to 1/10 of the original learning rate. We did not compare sudden drop with cosine drop....