DINO with low mAP in coco-like custom dataset

[letmejoin]

When use DINO or deformable-detr, my dataset has always low mAP, but cascade-rcnn-x101-32x4d, cascade swin-B/L has good performance. Is code of DINO has some bugs? Meanwhile, Co-DETR implacation with mmdet-2.25 with good accuracy.

[letmejoin]

has guys good idea?

[hhaAndroid]

@letmejoin Your comparison seems somewhat unfair. In my opinion, you can choose the same backbone for experimentation. However, I find your experiment to be very meaningful. Would it be possible to have a conversation to discuss it further?

[letmejoin]

@letmejoin Your comparison seems somewhat unfair. In my opinion, you can choose the same backbone for experimentation. However, I find your experiment to be very meaningful. Would it be possible to have a conversation to discuss it further?

cascade-x101 and DINO-x101 use same backbone

[hhaAndroid]

@letmejoin We will be conducting some studies in the near future. If you are interested, we can discuss it further. Alternatively, you can take a look at the Cascade DETR paper, which provides some analysis. However, I can assure you that the DINO algorithm in MMDetection is working fine, as I ran it yesterday.

[letmejoin]

@letmejoin We will be conducting some studies in the near future. If you are interested, we can discuss it further. Alternatively, you can take a look at the Cascade DETR paper, which provides some analysis. However, I can assure you that the DINO algorithm in MMDetection is working fine, as I ran it yesterday.

thanks for replay, i will read the paper for some information. BTW, Co-DINO-DETR work fine

[hhaAndroid]

@letmejoin Because Co-DINO-DETR incorporates both one-stage and two-stage auxiliary branches.

[RYHSmmc]

Same issue, @letmejoin , can you get a better result on coco-like dataset using Dino detecter? After my experiments, I doublt Dino can only get a SOTA result on Coco?