Description

This is a PyTorch Re-Implementation of CRAFT: Character Region Awareness for Text Detection.

Model	Recall	Precision	F-score
Original	93.1	97.4	95.2
Re-Implement	91.98	92.32	92.15

Prerequisites

Only tested on

• Anaconda3
• Python 3.7.1
• PyTorch 1.0.1
• opencv-python 4.0.0.21
• easydict 1.9

Installation

1. Clone the repo

git clone https://github.com/SakuraRiven/CRAFT.git
cd CRAFT

2. Data & Pre-Trained Model

• ICDAR2013, ICDAR2017, SynthText

• VGG16

• Pre-trained model

• Finetune model

Make a new folder pths and put the download pths into pths

mkdir pths
cd pths
mkdir backbone pretrain ft
cd ..
mv vgg16_bn-6c64b313.pth pths/backbone/
mv model_iter_50000.pth pths/pretrain/
mv model_iter_31600.pth pths/ft/

Here is an example:

.
├── CRAFT
│   ├── evaluate
│   ├── pths
│   └── sync_batchnorm
└── data
    ├── ICDAR2013
    ├── ICDAR2017
    └── SynthText

Train

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py

Finetune

CUDA_VISIBLE_DEVICES=0,1,2 python finetune.py

Detect

CUDA_VISIBLE_DEVICES=0 python detect.py

Evaluate

CUDA_VISIBLE_DEVICES=0 python eval.py