Closed zengwang430521 closed 2 years ago
The content parts of CDN queries are taken as ground-truth label embeddings with noises same as in DN-DETR.
Thanks for your reply.
Hi following up on this question,
I see that an indicator was used in DN-DETR but dropped in DINO-DETR. I quoted below the description of the indicator from the DN-DETR.
We also append an indicator to label embedding. The indicator is 1 if a query belongs to the denoising part and 0 otherwise.
So it was essentially used as the last dimension of the content query telling DETR if it is a denoised object or not.
Can you comment why you dropped the indicator? I couldn't find anywhere in the paper mentioning this. Also, I could not find ablation studies in DN-DETR showing that this indicator explicitly worked.
For context, I am trying to implement DN / potentially CDN in a DETR deriative.
Also very nice work!
Hi, I'm confused about what the content query of CDN queries is. Is it the same as the normal learnable content queries?
Looking forward to your reply. Thanks in advance!