-
Notifications
You must be signed in to change notification settings - Fork 63
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
release the implementation of adding BorderAlign Module to the RetinaNet? #15
Comments
I am very sorry that all of our experiments were done in another codebase. But we have only reproduced the primary experiment from the previous codebase without BorderAlign on RetinaNet. |
Thanks for your reply. There may be something wrong with what I understand. "we directly select the one with the highest score from the nine prediction boxes of each pixel to refine." I understand it to mean that the feature map in front of the BorderAlign of RetinaNet is the same as that of FCOS. Therefore, I did ArgMax on the coarse classification score first, got the highest one of the 9 anchors as the coarse bounding boxes, and kept the same operation as FCOS in the Border Align Module. The following is my implementation of BorderHead, please take a look and tell me what is wrong. Thank you very much.
|
retinanet.res50.800size.1x.tsd.zip |
Thanks for sharing. I still have three questions:
|
|
Thank you very much but I'm a bit confused about question 2. if the dimension of sec_reg_conv is [N, HW9, 256], how to reshape it to (N, -1, int(H * W))? |
I'm sorry that I reply to you so late.
|
Thanks for your excellent work and nice open-source code.
"Our BorderDet can be easily integrated with the many popular object detectors, e.g. RetinaNet and FPN.To prove the generalization of the BorderDet, we first add the proposed border alignment module to the RetinaNet. For a fair comparison, without modifying any setting of RetinaNet, we directly select the one with the highest score from the nine prediction boxes of each pixel to refine."
I tried to reproduce it as described in the paper, but the model has not converged well. so, can you please release the implementation of adding BorderAlign Module to the RetinaNet? thanks.
The text was updated successfully, but these errors were encountered: