-
Notifications
You must be signed in to change notification settings - Fork 9.5k
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
Fixed bugs about cascade_roi_head. #3244
Conversation
* Fixed the bug of refining bg rois at training when reg_class_agnostic is False. * Fixed the bug of potentially infering roi label to be bg at inference.
Hi @Jokoe66 , |
Hi, I don't validate it yet due to gpu limit.
Besides, I think the training logic of class-aware cascade roi head needs further modification. Under current modification, class-aware cascade roi head does not refine negative rois at training, but do it at inference by regarding them as their closest objects. I think it's better to keep the behavior at training consistent with that at inference. So I will change the labels of bg rois used to refine rois at training to what will be at inference.
|
The code looks good to me now. I will benchmark some models to validate the modification. |
Sure. Thanks. |
The performance seems OK. I ran cascade mask r50 three times and obtained 41.1/35.9, 41.2/35.9, and 41.1/35.8 box/mask mAP. |
* Fixed bugs about cascade_roi_head. * Fixed the bug of refining bg rois at training when reg_class_agnostic is False. * Fixed the bug of potentially infering roi label to be bg at inference. * Reformatted code. * Ensured foreground roi_labels when refining bboxes at training. * Fixed typo. * Fixed incompatibility between mmdet2.2(ops/conv_ws.py) and mmcv1.0.0 * Fixed typo.
Fixed the bug of refining bg rois at training when reg_class_agnostic is False.
Fixed the bug of potentially infering roi label to be bg at inference.