Hello friends,
I encountered a problem regarding the evaluation of semantic segmentation. Qualitatively the ground truth and prediction are quite similar but dice shows a small number about 0.56, How's this possible? If anyone has encountered the same problem so far, I would greatly appreciate it if you can share your experience and recommendation with me.
Hello friends,
To address this question, I think there are a few problems you are facing:
- First, about the dice loss formula that we are using. Usually I will use the formula quoted in the article (link: https://arxiv.org/pdf/1606.04797.pdf).
- Second, you can refer to the popular semantic segmentation architecture such as Attention U-Net or Res-Attention UNet.
Sincerely,
Hoang Nhut Huynh
1 .. use dice coefficient score
2 ..check ur object based F1 score, and mIoU if these are better then ur dice score implementation have a problem.
3 .. mostly dice score evaluate binary classes, check ur predicted masks if there are more then two values in the mask then you should check it on binary classes.
4 .. if all is good then may be Dice score is correct and ur eyes cannot see the difference.
Maryam Tavakol Elahi Take a look Article Unified Focal loss: Generalising Dice and cross entropy-base...
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts