saicoco / tf-sotd

Based EAST implements "Self-organized Text Detection with Minimal Post-processing via Border Learning"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

about training on icdar2015 dataset

happog opened this issue · comments

commented

Hi, did you train sotd on icdar dataset? I have trained it on icdar2015 but the predicted map is all background class. Should I use another loss function such as cross entropy? Thank you!

Sorry to I reply to so late, I didn't train it on icdar, the backgroud, border and text area are imbalance, so you need to add OHEM or other balanced loss function to train it, then it will work well