search

ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
767 stars 254 forks source link

No output boxes for custom training dataset. #146

Open ajinkya933 opened 4 years ago

ajinkya933 commented 4 years ago

I followed your code but when I retrain the model using python2 I coudnt reproduce your results, rather all the output files donot have any box coordinates,

Training batchsize=16 Platform: AWS p2xlarge (1 Nvidia GPU) python version: python2

I followed a discusion thread in issues which said "pixel_link-mingyu.guo" implementation is much better than yours and he has done converting your code to python3.

But in this case even that failed, heres his repo https://github.com/baby313/pixel_link

How can I get output by retraining?

File sizes of trained model are as follows:

model.ckpt-808.data-00000-of-00001   :    246MB
model.ckpt-808.index                            :     6.7KB
model.ckpt-808.meta                             :     1.5MB
mohamad-hoseini commented 4 years ago

i have the same problem, too. Training works on icdar2015 dataset, But doesn't work on my dataset. Not even a box detected after testing on a model at step 300.

Did You manage to solve it?

arshv27 commented 3 years ago

@ajinkya933 @mohamad-hoseini Were you guys able to resolve this issue?