search

strongwolf / o2f

One-to-Few Label Assignment for End-to-End Dense Detection (CVPR2023)
Apache License 2.0
40 stars 3 forks source link

Some questions about ambiguous anchor #1

Closed zkyseu closed 1 year ago

zkyseu commented 1 year ago

Hi Thanks for your wonderful work! I have a confusion about the ambiguous anchors. In o2f, you set the number of ambiguous anchors to 7. During the training process, the weight of the ambiguous anchors decreases gradually in classification loss. However, why the weight of ambiguous anchors does not decrease in GIOU loss? Do you try the experiments about decreasing the weight of ambiguous anchors in GIOU loss? Thanks for your replying!

strongwolf commented 1 year ago

In my opinion, weighting the regression branch is not helpful to dumplication removal beacuse the predictions are ranked by classfication scores during inference. However, I am not sure wheather weighting the regression branch in such a way as in the classificaion branch can improve the final performance. I will try the experiments soon.

zkyseu commented 1 year ago

Thanks! I look forward to your experimental results!

256-7421142 commented 10 months ago

Hi Thanks for your wonderful work! I have a confusion about the ambiguous anchors. In o2f, you set the number of ambiguous anchors to 7. During the training process, the weight of the ambiguous anchors decreases gradually in classification loss. However, why the weight of ambiguous anchors does not decrease in GIOU loss? Do you try the experiments about decreasing the weight of ambiguous anchors in GIOU loss? Thanks for your replying!

Hi,May I ask what the results of the experiment are now?

256-7421142 commented 10 months ago

In my opinion, weighting the regression branch is not helpful to dumplication removal beacuse the predictions are ranked by classfication scores during inference. However, I am not sure wheather weighting the regression branch in such a way as in the classificaion branch can improve the final performance. I will try the experiments soon. Hi,May I ask what the results of the experiment are now?