search

argman / EAST

A tensorflow implementation of EAST text detector
GNU General Public License v3.0
3.02k stars 1.05k forks source link

why dice loss? #68

Open BowieHsu opened 6 years ago

BowieHsu commented 6 years ago

Hi, I found loss function in model.py is dice classification loss which is different from the origin paper's balanced cross-entropy，does there exists paper provide the algorithm about dice classification? Thanks.

BYJRK commented 6 years ago

I have the same question. By the way, the Dice loss (1-2*IoU) looks kind of like the balancing factor beta in the class-balanced cross-entropy formula.

BowieHsu commented 6 years ago

@BYJRK I read this paper “Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations” which compared dice loss and Weighted cross-entropy, the paper result shows that dice loss on segmentation task maybe more robust, i tried dice loss on my own task and found it gets better performance.

BYJRK commented 6 years ago

@BowieHsu Thanks a lot! This is really helpful.