Connectionist Text Proposal Network in MXNet

Introduction

CTPN is a nice scene text detection method.

中文文档

Training

Execute script init.sh(init.bat on Windows) to initialize project.
Download pretrained model from here into model folder.
Download dataset from google drive or baidu yun. This dataset is already prepared by @eragonruan to fit CTPN.
Unzip the dataset downloaded to 'VOCdevkit' folder, and set both default.root_path and default.dataset_path in rcnn/config.py to '<somewhere>/VOCdevkit/VOC2007'. You can also change other hyperparams in rcnn/config.py.
Run python train_ctpn.py to train. Run python train_ctpn.py --gpus '0' --rpn_lr 0.01 --no_flip 0 to train model on gpu 0 with learning rate 0.01 and with flip data augmentation.

Testing

Use python demo_ctpn.py --image "<your_image_path>" --prefix model/rpn1 --epoch 8 to test.

Our results

NOTICE: all the photos used below are collected from the internet. If it affects you, please contact me to delete them.

Requirements: Hardware

Any NVIDIA GPUs with at least 2GB memory should be OK.

References

TODO

Custom dataset preparation tutorial
Windows support
Deploying network and c++ inference support