How to reproduce the performance of f1-score=80.83 on ICDAR2015? about east HOT 8 CLOSED

whats your performance ? I think the default configuration can reproduce that, in about 10000 steps 14*4gpu.

from east.

Thank you.

from east.

Hi, @argman , I found that the annotation data(gt_*.txt) icdar13 is different from icdar15.(icdar13 gt files only has 4 coodinates, and icdar 15 has 8 coods)
And your code in icdar.py seems like only support reading ground_truth file of icdar15 format. So can you tell me how do you train the EAST on both icdar15 and icdar13? (Did you generate the same type of annotation for icdar13 according to icdar15?)

Besides, how the code implements the "multiple scale testing" function?

Btw, I can only get the performance of around 77% by only training on icdar15 training set.

Thank you!

from east.

this is what i changed for the icdar13

tlx, tly, brx, bry = list(map(float, line[:4]))
x1, y1, x2, y2, x3, y3, x4, y4 = tlx, tly, brx, tly, brx, bry, tlx, bry

from east.

@tianzhuotao , thats cool, can you share your model or ideas ?

from east.

@argman , Hi, Could you please tell me the MS result on ICDAR2015 test set?

from east.

by improving the model a little bit, the f1-score can reach nearly 82 on ICDAR2015 test set (trained with ICDAR2013 AND ICDAR2015 training data)

can you show your idea? Thanks a lot! My email: [email protected]

from east.

by improving the model a little bit, the f1-score can reach nearly 82 on ICDAR2015 test set (trained with ICDAR2013 AND ICDAR2015 training data)

can you share your idea?Thank you!

from east.