Mmdetection: Add a new bbox regression loss
Recently the Generalized IoU loss has been proposed to replace the regression loss in detectors (https://arxiv.org/abs/1902.09630). Any plans to add it?
sovrasov
All 12 comments
Yes we plan to add it by the beginning of next month, but it is quite simple so you may also try it first.
hellock
on 14 Mar 2019
I've already implemented that, but my code rather looks like a hack, so I'm waiting for the clean implementation.
upd: you can check out my implementation here https://github.com/grib0ed0v/mmdetection/tree/vs/focal
sovrasov
on 15 Mar 2019
where did you add your regression loss into the project?I can not find the place.
Ezra-Yu
on 16 Mar 2019
Does it make sense? can you show me some experiments? Thx
zytx121
on 19 Apr 2019
For R-50 Faster R-CNN, GIoU loss is 1 points higher than the baseline, and IoU loss is 0.9 points higher. These implementations will be released after our refactoring.
hellock
on 19 Apr 2019
Amazing! I am much expecting!
zytx121
on 19 Apr 2019
where should i add your regression loss into the project? Is it in bbox_head.py?
gazelxu
on 19 Apr 2019
There are always something wrong when I implemented GIoU loss. So I'm waiting for the implementation or can you show me some experiments? Thx
gazelxu
on 13 May 2019
Any updates on when it will be added?
JulienSiems
on 10 Aug 2019
For R-50 Faster R-CNN, GIoU loss is 1 points higher than the baseline, and IoU loss is 0.9 points higher. These implementations will be released after our refactoring.
Hello @hellock @sovrasov I have implemented GIoU loss for libra faster rcnn. For VOC2007 test, I got 0.7points higher, while for COCO 2017val, I got 10points lower. I found that the acc increased slower than the baseline. Could you give me some advice? Waiting for your reply:)
DRosemei
on 29 Nov 2019
For R-50 Faster R-CNN, GIoU loss is 1 points higher than the baseline, and IoU loss is 0.9 points higher. These implementations will be released after our refactoring.
Have you tried cascade RCNN? I got a bad result.
ZYN-guy
on 16 Dec 2019
I've already implemented that, but my code rather looks like a hack, so I'm waiting for the clean implementation.
upd: you can check out my implementation here https://github.com/grib0ed0v/mmdetection/tree/vs/focal
too many values to unpack ? How to solve it? Than you
ZYN-guy
on 17 Dec 2019
Related issues
BeBeauty
路
3Comments
songyuc
路
3Comments
letanloc1998
路
3Comments
FrankXinqi
路
3Comments
fmassa
路
3Comments
Most helpful comment
For R-50 Faster R-CNN, GIoU loss is 1 points higher than the baseline, and IoU loss is 0.9 points higher. These implementations will be released after our refactoring.