search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.53k stars 11.68k forks source link

rpn_bbox_loss is too high, it's hard to lower #1687

Open zhaojw219 opened 5 years ago

zhaojw219 commented 5 years ago

When I train the network, rpn_bbox_loss is too high. It is difficult to reduce. Is there any good advice? Many thanks

Jingyixiong commented 4 years ago

Hellow zhaojw219. I also have the same problem as you, the rpn_bbox_loss is very high and I can not decrease its value. Have you already got some solutions for that?

ignaciovidalfranco commented 4 years ago

Well, I do have the same problem. Could you share some details about your dataset?

Jingyixiong commented 4 years ago

I am doing the same project as you were doing, the steel defect detection. And the for the number of masks per image, actually I give every discontinuous object of one image a single mask to represent it. Since mask rcnn is an instance detection neural network. I have some assumptions that why the accuracy of mask r-cnn is not that good, one maybe because of some of the defects in the image have very slender shape, as far as I know, mask r-cnn is not good at detecting such instance. Another possible reason is that the mask r-cnn is too complicated for this kind of task since we are only doing classifications for 4 classes of defects. So maybe simpler structure like U-net will perform better in such task. Also, we borrowed the weight from COCO dataset which is not designed for this kind of classification that is why it may be not appropriate to initialize weight by using COCO weight. If you have any good explanation for this problem, please tell me. Thanks dude!

ignaciovidalfranco notifications@github.com 于2019年10月17日周四 上午6:25写道：

Well, I do have the same problem. Could you share some details about your dataset?

  • What kind of images?
  • How many classes?
  • How many masks per image?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/matterport/Mask_RCNN/issues/1687?email_source=notifications&email_token=AKVGUEWVSTYD6TC4A7MGM6LQO6BE7A5CNFSM4ILG3GA2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBN7POQ#issuecomment-542898106, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKVGUEQ4WW4B7PBYUEBMCPLQO6BE7ANCNFSM4ILG3GAQ .

FTRobles commented 4 years ago

Hi, do you find a solution for your problem. I am ussing a diferent dataset than you but the behavior of the net is similar to yours,