search

Media-Smart / vedastr

A scene text recognition toolbox based on PyTorch
Apache License 2.0
534 stars 100 forks source link

what make model improve ? #29

Closed nightfuryyy closed 4 years ago

nightfuryyy commented 4 years ago

Thank you for awesome repo. I saw your accuracy model (tps_resnet_bilstm_attn) is 84.95. that is better than this model. What did you change ? Thank you so much.

ChaseMonsterAway commented 4 years ago

Hi, @nightfuryyy

In fact, the accuracy of tps_resnet_bilstm_attn of version 1.1.0 is 83.98 (the accuracy in is wrong ), and the model release by deep-text-recognition-benchmark is 83.897. 84.95 is a mistake becasue i didn't turn off the data filter, thus the accuracy of ic15 & cute80 are wrong.

In current version (2.0.0), the better performance of tps_resnet_bilstm_attn benifit from the data. The size of ST dataset we used is larger (about 200k, total 5.7M).