search

ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
768 stars 255 forks source link

When training on my datasets , the loss =0.00000 #32

Closed wanghuok closed 6 years ago

wanghuok commented 6 years ago

I convert my datasets format from voc to icdar2015. And then, i use icdar2015_to_tfrecord.py to convert them to tfrecord file. I use this file to train pixel_link, but loss always 0.000. What's wrong with it? Thanks.

wanghuok commented 6 years ago

When i convert voc to icdar2015, i set labels from '###' to anyother text, solved it.