yuliang-liu / bezier_curve_text_spotting Goto Github PK

Hi, Can you provide the trained model google drive link and a small demo of how to use to make inferences? It will be really wonderful to see its performance in practice.

Can bezier model recognize Chinese texts end2end?
If so, is there any result model published?

Hello, ABCNet is a meaningful work, can you give the codes about how to use the lexicon to obtain the results of text spotting? Thank you. @Yuliang-Liu

你好~~
想请教几点问题：

1. 我需要训练中文检测+识别模型，来进行证照文件的识别。这份代码是在pytorch框架下的，可以直接拿来训练模型么？
2. 如可以，是否有预训练模型呢，可以方便我节省一定的训练时间。
3. 证照文件中，同时含有水平文字、倾斜文本和环形文本，这个模型，可以自动进行相应的识别？？
4. 中文包含生僻字。。这种情况下，数据集貌似只能自己手动标注么，没办法合成？ ---ctw数据集貌似只有不到4K个汉字字符，无法满足使用要求。。
   恳请博主指点迷津~~

File "/maskrcnn_benchmark/solver/build.py", line 26, in make_optimizer
if cfg.DARTS_ON:
File "b/python3.6/site-packages/yacs/config.py", line 141, in getattr
raise AttributeError(name)
AttributeError: DARTS_ON

Hello,
Thanks for the great work, I have a question please, how to fine-tune the model to adopt different classes with bboxes and another language than English (if there is a tutorial that can explain that please share it, thank you in advance🙏

非常感谢您的分享，通过您的代码程序未跑通，是否能提供一下你的运行环境，我想尝试一下

@Yuliang-Liu @stan-haochen

Can ABCNet detect long text-lines instead of words?

@Yuliang-Liu @stan-haochen
Thank you for open-sourcing your hard work,

Regarding the ABCNet implementation in Detectron2, when do you expect releasing that.
Also, further documentation on how to train ABCNet.

Dear author,

I want to try to train your model with total text dataset. Can you please give me total_train.json file to do that.

Thank you

我配了三天死活配不好，各种报错，是否可以提供torch版本呢？另外，这个仓库里我看到已经有了maskrcnn_benchmark，是否就不用在install structions里安装maskrcnn_benchmark了呢？

作者你好，
在你们的ABCNet论文里面有一个对比warp图。然后(a)是说用TPS+STN做的，想问下这部分有代码吗？或者能详细介绍下解决思路吗？在论文里没有看到详细介绍，谢谢

I want to use python tools/tests/single_demo_bezier.py as README.md, but got error cannot import name 'BezierAlign'.
I have installed maskrcnn_benchmark successfully, but it seems like maskrcnn_benchmark doesn't contain BezierAlign module.

How can I solve this? @Yuliang-Liu

def test(scale=2):
image_size = (2560, 2560) # H x W
output_size = (256, 1024)

File "D:\ocr_demo\bezier_curve_text_spotting-master\maskrcnn_benchmark\layers\nms.py", line 8, in
nms = amp.float_function(_C.nms)
File "D:\Anaconda\envs\pytorch_gpu\lib\site-packages\yacs\config.py", line 141, in getattr
raise AttributeError(name)
AttributeError: nms

xmin,ymin是指哪四个顶点的坐标值？是弯曲文本区域的最小外接矩形么？还是指一条长边的四个控制点？

Thanks for your contribution.
How to train the model. And how to get the training data,
Thanks

@Yuliang-Liu
Tks for sharing such a great work
iam having a question about the cps: in single_demo_bezier.py, how to get those values, cause when i tried with different images, the network cant output the text scene correctly.
And when I tried to use vis_bezier.py, the results are not good as yours:

Which parameters should I change to get the same result as yours ?

谢谢你，请您给个联系方式，麻烦了

@Yuliang-Liu
Include icdar2015 dataloader for training and predicting

Thanks for your wonderful work,
When the Syntext-150k dataset will be available?
Have you trained your recognition part with special characters?

Thanks for your great work and sharing the code.
The text detection results are not available on your arxiv paper, so can you provide the text detection results(Recall/Precision/F-measure) on your datasets(CTW1500, Total-Text, etc.) if possible? Thanks a lot.

作者你好，
请问当时你们有试验backbone换为r18吗？ 我将你们官方的v2_attn_r50中的backbone替换为r18+FPN后，检测是正常的，但是识别全为0，但我直接用的是官方的代码（数据集换为我们自己的，r50+BiFPN是正常work的），所以不知道是不是backbone换小了就完全不行了呢？