Skip to content

Conversation

@implus
Copy link
Contributor

@implus implus commented Jun 22, 2020

  • add Generalized Focal loss (GFL), including QualityFocalLoss(QFL) and DistributionFocalLoss(DFL)
  • add GFL backbone, head, configs, and pretrained models and logs

Copy link
Contributor

@liaopeiyuan liaopeiyuan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking forward to this getting merged into the main branch!

implus and others added 9 commits June 28, 2020 02:21
modify variables  L, T, R, B   to  left, top, right, bottom
modify variables into meaningful names:
* pt -> scale_factor
* a -> pos   b -> pos_label
* disl -> dis_left  disr -> dis_right
* wl -> weight_left  wr -> weight_right
update formats
@hellock
Copy link
Member

hellock commented Jun 29, 2020

LGTM. Waiting for the benchmarking results.

@ZwwWayne
Copy link
Collaborator

LGTM. Waiting for the benchmarking results.

For now, I have finished training some models, which produce the same AP as reported.
gfl_r50_fpn_1x_coco.py: 40.2, gfl_r50_fpn_mstrain_2x_coco.py: 42.9.
So I think it is fine to merge this PR today and we can upload the complete benchmarked models in the next batch.

@hellock hellock merged commit 39e6257 into open-mmlab:master Jun 30, 2020
mike112223 pushed a commit to mike112223/mmdetection that referenced this pull request Aug 25, 2020
* add gfocal losses

* add gfl head, its related code, configs

* fix 2x config bug

* add most model and log in gfl README

* update gfl model and log

* Update transforms.py

modify variables  L, T, R, B   to  left, top, right, bottom

* Update gfocal_loss.py

modify variables into meaningful names:
* pt -> scale_factor
* a -> pos   b -> pos_label
* disl -> dis_left  disr -> dis_right
* wl -> weight_left  wr -> weight_right

* Update gfocal_loss.py

update formats

* remove duplicated get_bboxes/IoU calculation, make QFL consistent

* add all docstring in gfl_head.py and gfocal_loss.py
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants