Skip to content

A PyTorch implementation of ECCV2018 Paper: TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes

License

Notifications You must be signed in to change notification settings

princewang1994/TextSnake.pytorch

Repository files navigation

TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes

A PyTorch implement of TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes (ECCV 2018) by Megvii

Paper

Comparison of different representations for text instances. (a) Axis-aligned rectangle. (b) Rotated rectangle. (c) Quadrangle. (d) TextSnake. Obviously, the proposed TextSnake representation is able to effectively and precisely describe the geometric properties, such as location, scale, and bending of curved text with perspective distortion, while the other representations (axis-aligned rectangle, rotated rectangle or quadrangle) struggle with giving accurate predictions in such cases.

Textsnake elements:

  • center point
  • tangent line
  • text region

Description

Generally, this code has following features:

  1. include complete training and inference code
  2. pure python version without extra compiling
  3. compatible with laste PyTorch version (write with pytroch 0.4.0)
  4. support TotalText and SynthText dataset

Getting Started

This repo includes the training code and inference demo of TextSnake, training and infercence can be simplely run with a few code.

Prerequisites

To run this repo successfully, it is highly recommanded with:

  • Linux (Ubuntu 16.04)
  • Python3.6
  • Anaconda3
  • NVIDIA GPU(with 8G or larger GPU memory for training, 2G for inference)

(I haven't test it on other Python version.)

  1. clone this repository
git clone https://github.com/princewang1994/TextSnake.pytorch.git
  1. python package can be installed with pip
$ cd $TEXTSNAKE_ROOT
$ pip install -r requirements.txt

Data preparation

Pretraining with SynthText

$ CUDA_VISIBLE_DEVICES=$GPUID python train.py synthtext_pretrain --dataset synth-text --viz --max_epoch 1 --batch_size 8

Training

Training model with given experiment name $EXPNAME

training from scratch:

$ EXPNAME=example
$ CUDA_VISIBLE_DEVICES=$GPUID python train.py $EXPNAME --viz

training with pretrained model(improved performance much)

$ EXPNAME=example
$ CUDA_VISIBLE_DEVICES=$GPUID python train.py example --viz --batch_size 8 --resume save/synthtext_pretrain/textsnake_vgg_0.pth

options:

  • exp_name: experiment name, used to identify different training processes
  • --viz: visualization toggle, output pictures are saved to ./vis by default

other options can be show by run python train.py -h

Running tests

Runing following command can generate demo on TotalText dataset (300 pictures), the result are save to ./vis by default

$ EXPNAME=example
$ CUDA_VISIBLE_DEVICES=$GPUID python eval_textsnake.py $EXPNAME --checkepoch 190

options:

  • exp_name: experiment name, used to identify different training process

other options can be show by run python train.py -h

Evaluation

Total-Text metric is included in dataset/total_text/Evaluation_Protocol/Python_scripts/Deteval.py, you should first modify the input_dir in Deteval.py and run following command for computing DetEval:

$ python dataset/total_text/Evaluation_Protocol/Python_scripts/Deteval.py $EXPNAME --tr 0.8 --tp 0.4

or

$ python dataset/total_text/Evaluation_Protocol/Python_scripts/Deteval.py $EXPNAME --tr 0.7 --tp 0.6

it will output metrics reports.

Pretrained Models

Download from links above and place pth file to the corresponding path(save/XXX/textsnake_vgg_XX.pth).

Performance

DetEval reporting

Following table reports DetEval metrics when we set vgg as the backbone(can be reproduced by using pertained model in Pretrained Model section):

tr=0.7 / tp=0.6(P|R|F1) tr=0.8 / tp=0.4(P|R|F1) FPS(On single 1080Ti)
expand / no merge 0.652 | 0.549 | 0.596 0.874 | 0.711 | 0.784 12.07
expand / merge 0.698 | 0.578 | 0.633 0.859 | 0.660 | 0.746 8.38
no expand / no merge 0.753 | 0.693 | 0.722 0.695 | 0.628 | 0.660 9.94
no expand / merge 0.747 | 0.677 | 0.710 0.691 | 0.602 | 0.643 11.05
reported on paper - 0.827 | 0.745 | 0.784

* expand denotes expanding radius by 0.3 times while post-processing

* merge denotes that merging overlapped instance while post-processing

Pure Inference

You can also run prediction on your own dataset without annotations:

  1. Download pretrained model and place .pth file to save/pretrained/textsnake_vgg_180.pth
  2. Run pure inference script as following:
$ EXPNAME=pretrained
$ CUDA_VISIBLE_DEVICES=$GPUID python demo.py $EXPNAME --checkepoch 180 --img_root /path/to/image

predicted result will be saved in output/$EXPNAME and visualization in vis/${EXPNAME}_deploy

Qualitative results

  • left: prediction/ground true
  • middle: text region(TR)
  • right: text center line(TCL)

What is comming

  • Pretraining with SynthText
  • Metric computing
  • Pretrained model upload
  • Pure inference script
  • More dataset suport: [ICDAR15, CTW1500]
  • Various backbone experiments

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgement

About

A PyTorch implementation of ECCV2018 Paper: TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published