search

ayumiymk / aster.pytorch

ASTER in Pytorch
MIT License
663 stars 169 forks source link

My accuracy of SVTP of lexicon0 is 80.9 rather than the ReadMe's 81.2, is there somewrong? #22

Closed sonack closed 4 years ago

sonack commented 4 years ago

image Is there something wrong?

sonack commented 4 years ago

FYI, Below is my tested your pretrained models' performace:

IIIT5k SVT IC03 IC13 IC15 SVTP CUTE
ASTER (L2R) 92.67 - 93.72 90.74 - 78.76 76.39
ASTER.Pytorch 93.2 89.2 92.2 91 78.0 81.2 81.9
ASTER.Pretrain 93.2 89.2 92.8 93.3 78.0 80.9 81.9
ayumiymk commented 4 years ago

emmm... Except for SVTP, you can obtain the same even better results on other datasets. Maybe the images of SVTP I used is a little different from you. But I think 0.3 should be ok. If I have time, I will check that.

sonack commented 4 years ago

emmm... Except for SVTP, you can obtain the same even better results on other datasets. Maybe the images of SVTP I used is a little different from you. But I think 0.3 should be ok. If I have time, I will check that.

I used the tensorflow version of Aster's released test datasets https://github.com/bgshih/aster/releases/, so may it be a little different from yours?

ayumiymk commented 4 years ago

I can guarantee that the original images in both the tensorflow and pytorch versions are the same. However, there may be some different image processings to generate tfrecoder and lmdb. You can check the create_xxx.py to find the reasons.

sonack commented 4 years ago

I can guarantee that the original images in both the tensorflow and pytorch versions are the same. However, there may be some different image processings to generate tfrecoder and lmdb. You can check the create_xxx.py to find the reasons.

thanks, I just replace the proper file to be gt.txt and use your https://github.com/ayumiymk/aster.pytorch/blob/master/lib/tools/create_svtp_lmdb.py file.

shaohailin commented 4 years ago

@sonack icdar2013 in the paper has 1015 images,but the dataset in https://github.com/bgshih/aster/releases/ only has 857 images i use the 857 images of icdar2013 and test it ,the result is 93.3% @ayumiymk could you sent the link of the icdar2013_1015 dataset and it's gt.txt in the paper ,