search

JialeCao001 / D2Det

D2Det: Towards High Quality Object Detection and Instance Segmentation (CVPR2020)
https://openaccess.thecvf.com/content_CVPR_2020/papers/Cao_D2Det_Towards_High_Quality_Object_Detection_and_Instance_Segmentation_CVPR_2020_paper.pdf
MIT License
297 stars 86 forks source link

seems to a bug? #37

Open clelouch opened 3 years ago

clelouch commented 3 years ago

image it seems that the mask_scores = 0.3 * mask_iou + bbox_score

clelouch commented 3 years ago

image if i am right, it should be "if label is not None"?

JialeCao001 commented 3 years ago

@clelouch Thanks for interest. I am sorry that I donot clean the code very well. It is 0.3 * mask_iou + bbox_score. We donot use the original multiplication in mask scoring paper. We always implement the code in L344.

clelouch commented 3 years ago

Thanks for your reply. I am curious

  1. how your determine the value of the hyperparameter (i.e. 0.3 ) and
  2. compared with the original multiplication in mask scroing rcnn, how much difference it makes .
  3. Besides, why you choose such a way to generate the mask score?
  4. is the D2Det sensitive to the hyperparameter(i.e. 0.3 )? looking forward to your reply!