problem in the output about east HOT 6 CLOSED

@rmmal , thanks for trying EAST, this is interesting,
I think you can get reasonable result training until 7000+ iterations.
can you provide a sample annotation of your data ?

In #18 , the result seems nice.

from east.

You are welcome @argman , Great effort btw.

As i said before i had a x1,y1 and x2,y2
where i converted them to x1,y1, x2,y1, x2,y2, x1,y2

so for example the annotations of the last image is

550,1667,833,1667,833,1737,550,1737,text
213,1679,493,1679,493,1731,213,1731,text
230,1741,802,1741,802,1777,230,1777,text
357,1330,687,1330,687,1458,357,1458,text
253,1492,790,1492,790,1539,253,1539,text
159,984,864,984,864,1093,159,1093,text

what could be wrong ? suggestions ?

from east.

setting text-scale 1024 , and input size 512 could make such a problem ?

from east.

@rmmal , I think your problem is different. The definition of text maybe confusing for the network to learn. In the paper, text is defined as a single line. From your annotation, i cannot find what's the standard of a text line.
e.g. In the last image, why RADO TRUE AUTOMATIC DIAMONDS and HIGH-TECH forms a single block but not two lines

from east.

I edited the dataset and started now training again , I will see what will happen and tell you.

Thanks for your help @argman

from east.

@rmmal any updates regarding your training & results?

from east.