search

IDEA-Research / DAB-DETR

[ICLR 2022] Official implementation of the paper "DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR"
Apache License 2.0
501 stars 86 forks source link

question about two_satge: #43

Closed leayz-888 closed 2 years ago

leayz-888 commented 2 years ago

Thank you for your excellent work, What I want to know is have you used the two-stage strategy when training DAB-Deformable-DETR? For DAB-Deformable-DETR, does it give a performance boost?

SlongLiu commented 2 years ago

I believe you will get better results with a two-stage strategy, but we did not explore this in this paper.

leayz-888 commented 2 years ago

I believe you will get better results with a two-stage strategy, but we did not explore this in this paper.

Thanks for your reply, when I used the two-stage strategy to train my own dataset, I got a much worse result, AP dropped by 4%, the backbone I used was resnext101 (without using pretrained model). When I use the pre-trained model you provided for fine-tuning, AP drops by 2% compared to Deformable-DETR May I ask why this happens? Looking forward to your reply.

SlongLiu commented 2 years ago

I speculate that there might be some bugs in your implementation. We observe consistent improvements when introducing a two-stage strategy.