EATEN: Entity-aware Attention for Single Shot Visual Text Extraction

Accepted to ICDAR 2019 arxiv
Authors: He Guo, Xiameng Qin, Jiaming Liu, Junyu Han, Jingtuo Liu and Errui Ding

Abstract

This repository is designed to provide an open-source dataset for Visual Text Extraction.

Samples

Train ticket

Real images

Synthetic images

Some clean images

Some hard images

Passport

Some images

Some hard images

Business card

Downloads

The dataset can be downloaded through the following link:
baiduyun, PASSWORD: e4z1

Some details:

scenes	number	size	Google Drive link
train ticket	300k synth + 1.9 real	13G	dataset_trainticket.tar
passport	100k synth	5.8G	dataset_passport.tar
business card	200k synth	19G	dataset_business.tar.0 dataset_business.tar.1 dataset_business.tar.2 dataset_business.tar.3

Limitations&&Todo

• [A large of training data]
  Todo:
  1. Use CycleGan or domain adaptation to synth data to train EATEN.
  2. Introduce datasets of STR to EATEN.
• [Generalization on complex scenes]
  Todo:
  1. Add bounding box annotations of ToIs to EATEN, such as 2019-ICCV-oral Towards Unconstrained End-to-End Text Spotting.
• [Engineering]
  1. Merge server decoder to one.
  2. parallel decoding.