opconty / keras_std_plus_plus

This repository is the official implementation of `A Semantic-based Arbitrarily-Oriented Scene Text Detector`(named STD++ as it is the improved version of STD).

STD++: A Semantic-based Arbitrarily-Oriented Scene Text Detector

This repository is the official implementation of A Semantic-based Arbitrarily-Oriented Scene Text Detector(named STD++ as it is the improved version of STD).due to lack of computing resources and time, we tested STD++ on MTWI2018 dataset, and we hope to perform more experiments on any other benchmark datasets, such as IC15,IC17,COCO-Text,MSRA-TD500 and so on.

images come from icdar2017rctw

Introduction

STD++ is the improved version of STD, which solved STD's limitations and can be used to detect arbitrarily-oriented texts, yet still preserves its accuracy and efficiency:

• no any further post-processings, like NMS.
• anchor-free.
• easy to generate training labels.
• only one step process to get final bounding boxes.

Any questions or suggestions,please drop a comment or contact me,email: [email protected].

Training

Download RCTW17 dataset below, and configure your local directory path. refer to train.py

Inference

predict.py

Examples

Dataset

We trained STD++ on MTWI2018 dataset, training and testing images can be downloaded from this site for Text Localization, and we make STD++ annotations available on baiduyun, code: nuti.

License

This project is released under the Apache 2.0 license.

Citation

If you use our codebase in your research, please cite this project. a paper or technical report will be released soon.

And besides, you are welcomed to join us to maintain this project.

@misc{std_plus_plus2019,
  author =       {Gao Lijun},
  title =        {STD++: A Semantic-based Arbitrarily-Oriented Scene Text Detector},
  howpublished = {\url{https://github.com/opconty/keras_std_plus_plus}},
  year =         {2019}
}

keras_std_plus_plus's Issues

pretrained model

@opconty Thank you for your hard work,
Will you release a pretrained model so that we can test your code.

Also, how to use the model to detect an image?

paper

hi, it is has paper?

cannot import name 'Deconv2D'

Which version of tensorflow should i use?
When running train.py

Traceback (most recent call last):
  File "train.py", line 11, in <module>
    from models.losses import cls_center, regr_h, regr_offset, cl,angle
  File "/home/home/p13/keras_std_plus_plus/models/__init__.py", line 7, in <module>
    from .std_vgg16_skew import StdVGG16
  File "/home/home/p13/keras_std_plus_plus/models/std_vgg16_skew.py", line 10, in <module>
    from keras.layers import Input,Concatenate, Conv2D,UpSampling2D, BatchNormalization, Deconv2D
ImportError: cannot import name 'Deconv2D'

Training the mtwi dataset

@opconty @zhouyuangan

• Which version of tensorflow should i use?
• What changes did you make to config.py and train.py in-order to train the mtwi dataset?
• For training mtwi dataset, what is your folder structure?
• upload your trained model so i can test.

Inference Speed

@opconty As the title says, could you provide inference speed for me? I trained model on the mtwi dataset and ran predict.py to visualize result but cost 5 seconds per image, it's weird and confusing.

Generate XML annotations on own dataset

@opconty I have a dataset contains some gt files formated like x1,y1,x2,y2,...,x4,y4, label, I inspect your provided xml and confused me, could you provide script or code block for me? Thanks a lot.

can you share the pretrained model

Training icdar 2017

@opconty

How to train on icdar 2017?
I have noticed that you std++ code has it's own reconstructed annotations, so how can I prepare the icdar annotations and convert x1,y1,x2,y2,x3,y3 to std++ annotations?