ContourNet

This is a pytorch-based implementation for paper ContourNet (CVPR2020). ContourNet is a contour-based text detector which represents text region with a set of contour points. This repository is built on the pytorch maskrcnn.

ToDo List

• [x] Release code
• [x] Document for Installation
• [x] Trained models
• [x] Document for testing and training
• [x] Evaluation
• [x] Experiment on more datasets
• [ ] re-organize and clean the parameters

Updates

2020/5/6 We upload the models on Drive.
2020/6/11 We update the experiment for CTW-1500 and further detail some training settings.

Requirements

We recommend you to use Anaconda BaiduYun(passward:1y3v) or Drive to manage your libraries.

Step-by-step install

  conda create --name ContourNet
  conda activate ContourNet
  conda install ipython
  pip install ninja yacs cython matplotlib tqdm scipy shapely networkx pandas
  conda install pytorch=1.0 torchvision=0.2 cudatoolkit=9.0 -c pytorch
  conda install -c menpo opencv
  export INSTALL_DIR=$PWDcd $INSTALL_DIR
  git clone https://github.com/cocodataset/cocoapi.git
  cd cocoapi/PythonAPI
  python setup.py build_ext install
  cd $INSTALL_DIR
  git clone https://github.com/wangyuxin87/ContourNet.git
  cd ContourNet
  python setup.py build develop

Results

We use only official training images to train our model.

Experiment on IC15 dataset

Data preparing

step 1

Prepare data follow COCO format or you can download our IC15dataset BAIDU (passward:ect5) or Geogle Drive, and unzip it in

   datasets/.

step 2

You need to modify maskrcnn_benchmark/config/paths_catalog.py to point to the location where your dataset is stored.

step 3

Download ResNet50 model BAIDU(passward:edt8) or Drive and put it in ContourNet/.

Test IC15

Test with our proposed model BAIDU(password:g49o) or Drive

Put the folder in

   output/.

Set the resolution to 1200x2000 in maskrcnn_benchmark/data/transformstransforms.py (line 50 to 52). You can ignore this step when you train your own model, which seems to obtain better results. Then run

   bash test_contour.sh

Evaluate

Put bo.json to ic15_evaluate/, then run

   cd ic15_evaluate
   conda deactivate
   pip2 install polygon2
   conda install zip
   python2 eval_ic15

Train our model on IC15

As mentioned in our paper, we only use offical training images to train our model, data augmentation includes random crop, rotate etc. There are 2 strategies to initialize the parameters in the backbone:1) use the ResNet50 model (ImageNet)BAIDU(passward:edt8) or Drive, this is provided by Yuliang, which is ONLY an ImageNet Model With a few iterations on ic15 training data for a stable initialization.2) Use model only pre-trained on ImageNet(modify the WEIGHT to catalog://ImageNetPretrained/MSRA/R-50 in config/ic15/r50_baseline.yaml). In this repository, we use the first one to train the model on this dataset.

Step 1:

Run

   bash train_contour.sh

Step 2:

Change the ROTATE_PROB_TRAIN to 0.3 and ROTATE_DEGREE to 10 in config/ic15/r50_baseline.yaml (corresponding modification also needs to be done in maskrcnn_benchmark/data/transformstransforms.py from line 312 to 317), then finetune the model for more 10500 steps (lr starts from 2.5e-4 and dot 0.1 when step = 5k,10k).

Experiment on CTW dataset

Data preparing

step 1

Prepare data follow COCO format or you can download our CTW-dataset, and unzip it in

   output/.

step 2

You need to modify maskrcnn_benchmark/config/paths_catalog.py to point to the location where your dataset is stored.

Test CTW

Test with our proposed model Drive

Put the folder in

   output/.

Then run

   bash test_contour.sh

Evaluate

Run

   cd ctw_eval
   python eval_ctw1500.py

Train our model on CTW

Run

   bash train_contour.sh

Improvement

1. We use different reconstruction algorithm to rebuild text region from contour points for curved text, you can reproduce our approach used in the paper by modifying the hyper-parameter in Alpha-Shape Algorithm (some tricks also should be added). Furthermore, more robust reconstruction algorithm may obtain better results.
2. The detection results are not accurate when the proposal contains more than one text, because of that the strong response will be obtained in both contour regions of texts.
3. Some morphological algorithms can make the contour line more smooth.
4. More tricks like deformable_conv, deformable_pooling in the box_head, etc. can further improve the detection results.

Citation

If you find our method useful for your reserach, please cite

 @inproceedings{wang2020contournet,
  title={ContourNet: Taking a Further Step toward Accurate Arbitrary-shaped Scene Text Detection},
  author={Wang, Yuxin and Xie, Hongtao and Zha, Zheng-Jun and Xing, Mengting and Fu, Zilong and Zhang, Yongdong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11753--11762},
  year={2020}
  }

Feedback

Suggestions and discussions are greatly welcome. Please contact the authors by sending email to [email protected]

GitHub