If there any method to merge oriented small box into oriented line?

[dajiangxiaoyan]

For example, EAST default result is：

After merging:

[loulan-D]

Have you solved the problem about merging oriented small box into oriented line? I have some ideas about dealing with the problem . But i still have some questions . We can just chat with each other to implement the idea.

[dajiangxiaoyan]

@loulan-D Currently, I merge box following CTPN poly.

[loulan-D]

@dajiangxiaoyan Thank you for your reply. I want to know how about the effect? Have you run the EAST-demo by your personal dataset? I run the demo by personal dataset. There are 10000 pictures that have the correspondent　txt-file label The oriented small box doesn't appear by the obtained model . But there a new problem. The result is bad about the long text . As shown below.

About the long text . the ideal effect.

[codecolony]

I'm facing same issue too :(

[dajiangxiaoyan]

@loulan-D I have re-trained east by my own dataset. But still face the same problem. The long text images are bad. But EAST detect text angle is very accurate. For now, I use two model EAST and CTPN. And the policy is:

1. Use EAST model to detect text angle.
2. If most of text is the same angle, then rotate image and re-detect by CTPN model
3. Else still detect by CTPN model. Then save EAST skew box and CTPN horizontal box.

Like this. The horizontal box is detected by CTPN and the other one is by EAST

[dajiangxiaoyan]

@loulan-D Maybe east works bad on long-text for this: https://github.com/argman/EAST/issues/94

[dajiangxiaoyan]

@loulan-D If most of your images are horizontal boxes，maybe you should use CTPN.

[loulan-D]

@dajiangxiaoyan Thank you for your reply. I'm very interested in your results. I want to know how to choose score threshold and nms threshold about the only east used.and what are you result about the threshold? Meanwhile, many of my images are skew boxes, so i don't use the CTPN. I also saw the CTPN's paper, but i can't combine two advantages about the ctpn and east. I will have a try following your policy. Thank you again.

[dajiangxiaoyan]

@loulan-D I use the default score threshold and nms threshold in this https://github.com/argman/EAST/blob/master/eval.py#L69. But I get a better result when set max_side_len = 512 in https://github.com/argman/EAST/blob/master/eval.py#L39

[xylcbd]

nice work @dajiangxiaoyan

[kewin1807]

@dajiangxiaoyan can you share code when you merge box, thanks