Skip to content
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

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


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


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.


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
  1. python package can be installed with pip
$ pip install -r requirements.txt

Data preparation

Pretraining with SynthText

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


Training model with given experiment name $EXPNAME

training from scratch:

$ EXPNAME=example

training with pretrained model(improved performance much)

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


  • 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 -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 $EXPNAME --checkepoch 190


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

other options can be show by run python -h


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

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


$ python dataset/total_text/Evaluation_Protocol/Python_scripts/ $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).


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 $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


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



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




No releases published


No packages published