search

argman / EAST

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

Result is not so good #209

Open jianchong-chen opened 5 years ago

jianchong-chen commented 5 years ago

I run the demo with pre-trained model, and I find that the result is not as good as the paper descriped. Even through there are only numbers and english letters. Should I re-train the model with my dataset? output

kapness commented 5 years ago

you can use another repo Advanced_East to solve this problem, paper points out that east can't process long text line well@jianchong-chen

renukachittimalla commented 5 years ago

Hello,

I have the same problem too. the model performance is not consistent. I have tested the model to understand the limits of model on long texts with varying font size. Please suggest if there could be any changes made. Also can you please suggest if changing receptice field size could improve the performance? img_22 img_4

rogersct commented 5 years ago

Hello,

I have the same problem too. the model performance is not consistent. I have tested the model to understand the limits of model on long texts with varying font size. Please suggest if there could be any changes made. Also can you please suggest if changing receptice field size could improve the performance? img_22 img_4

May i know where to find this receptive field size parameter?

austingg commented 5 years ago

@rogersct receptive field size can be calculate based on your network parameter such as conv kernel size, stride.

jianchong-chen commented 5 years ago

@kapness thanks for your help! I will try it later.

daniiki commented 3 years ago

@jianchong-chen Have you tried changing network parameters? Does it work?