exuse me！！！！i have question,my data is image which full of long words,just like books,but when i train model by my data,it can not detect all the words in one line.maybe it can only detect half of a line and my input_size is 768 and text_scale is 768.could you tell me where i maybe wrong?thank you so much

[cjt222]

exuse me！！！！i have question,my data is image which full of long words,but when i train model by my data,it can not detect all the words in one line.maybe it can only detect half of a line and my input_size is 768 and text_scale is 768.could you tell me where i maybe wrong?thank you so much @argman @zxytim

[argman]

.............can you provide some sample images?

[cjt222]

i am sorry , i ever thought issue can not show image.......@argman just like this image

[argman]

currently EAST is more suitable for detecting words(or short text), if you want to detect long texts, you can resize the width side. If your images contain all like this, I will recommend CTPN

[cjt222]

could you tell me why it can not detect long text? in fact, is long text regression hard to train? @argman

[argman]

no, because the receptive field of CNN is too small

[cjt222]

yes, i see, thank you so much!!!

[xiaomaxiao]

@cjt222 I also face such a problem, do you have a solution?

[ravikantb]

@xiaomaxiao , @cjt222 : Did you guys manage to solve it? I am also stuck here. In case I resolve it before you guys revert then will update it here.

[xiaomaxiao]

@ravikantb now , i use ctpn

[ravikantb]

@xiaomaxiao : Thanks for your response. Could you please tell me if you are using following implementation of CTPN: https://github.com/eragonruan/text-detection-ctpn or you have implemented your own?

[xiaomaxiao]

@ravikantb yes , we can talk together 。

[ravikantb]

Sure, thanks! :)

[dajiangxiaoyan]

CTPN works badly on oriented images. Does anyone make east to work on horizontal long text images?

[gangliu123]

HAHA， same issue. I am looking for the same solution as well @dajiangxiaoyan

[codecolony]

Let me know if anyone has an alternative or solution for this. CTPN doesn't fit for my needs.

[codecolony]

@argman - You mentioned that the receptive field of the CNN is too small. Does tweaking this CNN part be of any use? Did anyone try it?

[renukachittimalla]

@argman I have faced the same problem with long text too. But the main problem is the model performance is not consistent. I have done the test to understand the limits of the model on long text with varying fontsizes. Sometime it captures the most of the test and sometimes it misses the same whole line which it has already detected previously in another line.PFB the images for better understanding. Also the resnet-50 has full image receptive field so how is it going to effect to detect long text? If effecting, did anyone try modifying the receptive field to make the model work in long text?

thanks in advance

[wiistriker]

@renukachittimalla did you solve this issue?

[ghost]

>>>frozen_model.pb<<<

[wiistriker]

@DonkeySmall what is this? how did you train that model?

[ghost]

This is EAST model trained on my own dataset, I don't remember the training parameters, and I don't have the dataset left.

[ghost]

Something like this

[daniiki]

hey has anyone solved the problem?? :) I really need a solution for this problem

[wiistriker]

@argman @zxytim @renukachittimalla is there any solution yet? EAST cant detect long text lines?

[khu834]

What sort of aspect ratio of "long text" are you looking to detect?

The two main knobs for optimizing performance on long text is:

1. resolution (essentially number of pixels each letter occupies)
2. training data (if your training data represents the types of text aspect ratio you are expecting)

I'd suggest tuning the model using (1) and (2) above.