TFDBNet

A TensorFlow 2 reimplementation of Real-time Scene Text Detection with Differentiable Binarization available as a Python package and using TedEval for evaluation metrics.

Data Preparation

Store images in imgs folder and groundtruths in gts folder. Then, prepare text files for training and validate data in the following format with '\t' as a separator:

Example for ICDAR 2015 train.txt:

./datasets/train/train_imgs/img_1.jpg	./datasets/train/train_gts/gt_img_1.txt
./datasets/train/train_imgs/img_2.jpg	./datasets/train/train_gts/gt_img_2.txt

Example for ICDAR 2015 validate.txt:

./datasets/validate/validate_imgs/img_1.jpg	./datasets/validate/validate_gts/gt_img_1.txt
./datasets/validate/validate_imgs/img_2.jpg	./datasets/validate/validate_gts/gt_img_2.txt

You can customize the script in dir2paths.sh to generate the above train.txt and validate.txt for your own dataset. And the groundtruths can be .txt files, with the following format:

x1,y1,x2,y2,x3,y3,x4,y4,annotation

Below is the content of ./datasets/train/train_gts/gt_img_1.txt:

377,117,463,117,465,130,378,130,Genaxis Theatre
493,115,519,115,519,131,493,131,[06]
374,155,409,155,409,170,374,170,###
492,151,551,151,551,170,492,170,62-03
376,198,422,198,422,212,376,212,Carpark
494,190,539,189,539,205,494,206,###
374,1,494,0,492,85,372,86,###

Quick Start

pip install tfdbnet

After installation, see the demo on ICDAR 2015 dataset to know how to use. You can download my example trained weights along with the 2 files train.txt and validate.txt mentioned above here.