search

argman / EAST

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

python eval.py get pool results ? #96

Closed ccnankai closed 6 years ago

ccnankai commented 6 years ago

hi @argman I use icdar2017 data to train, based on your results on icdar2015 and 2013（model.ckpt-49491）, but why the test results are zero? 0 text boxes before nms ? This problem has troubled me for a long time, Can you help me solve this problem? hankyou very much .

image

ccnankai commented 6 years ago

@zxytim @QaisarRajput

QaisarRajput commented 6 years ago

This can be anything provide more info i.e sample of input data. but the first thing i can suggest is to look at the tensorboard of the trained model. more precisely the Images. (score_maps, geo_maps etc) those will get you an idea whether the model you trained is fine or not. 0 text boxes are weird..

argman commented 6 years ago

yes, I write all the training details in tensorboard

JasonDoingGreat commented 6 years ago

@argman @QaisarRajput I also tried to fine-tune the 49491 model on icdar 2017 chinese dataset, which has about 8000 training images, but after nearly 5000 steps of training, the total loss does not converge and dumping around 0.015. Then I tried to fine-tune from resnet model with the same dataset. Finally got the same dumping thing. Tensorboard details are listed below. image

image

I was wondering what went wrong here since the training does not converge and test results are poor?

SivaKanishka commented 6 years ago

@ccnankai Could you resolve the "0 text box before nms" error