2019-03-01Back to list
The Competitions of International Conference on Document Analysis and Recognition (ICDAR), also known as the "World Cup" of document analysis and recognition, has begun this year on March 1st.
As one of the main organizers of the ICDAR 2019 Robust Reading Competitions, Baiducollaborated with the University of Malaya, South China University of Technology, Computer Vision Centre in Spain, and China Society of Image & Graphics to organize two challenges: ICDAR 2019-LSVT(Large-scale Street View Text with Partial Labeling) and ICDAR 2019-ArT(Arbitrary-Shaped Text).
The prize money for the challenges totals at US$17,400 and it is open to domestic and foreign academic institutions, companies, and other organizations. The awards ceremony will be held during the ICDAR in September 2019.
Introduction to the Competitions
ICDAR is the flagship global conference for the recognition of text and image analysis, and it has been taking place biennially since 1991.
Since ICDAR established the "Robust Reading Competitions" in 2003, the competitions have become an important international standard for evaluating and examining the latest technological development for automatic detection and smart recognition of textual information in online images or natural images of real-world situations. Many methods proposed in the Competitions have made tremendous contributions to the advancements of OCR technologies. Today, more than 3,500 teams from 89 countries take part in the competitions.
From 2015 to 2018, Baidu reaped multiple number one rankings in the challenges of ICDAR 2013 and ICDAR 2015 for text detection, recognition and end-to-end tasks, retaining its leading position in the industry.
Baidu applied deep learning-based OCR technology to multiple business scenarios and created universal text recognition capabilities for documents, translation, online images, street view and more. Such services are provided in ten languages including Chinese, English, Japanese, Korean, Portuguese, German, French, Italian, Spanish, and Russian.
Baidu has various OCR product lines for key applications in finance, insurance, and transportation such as solutions to recognize certificates and receipts. Baidu’s OCR models are used in many smart devices and its SDK empowers multiple mobile phone manufacturers to create leading visual computing capabilities for end users.
ICDAR2019 Robust Reading Challenge on Large-scale Street View Text with Partial Labeling (ICDAR 2019-LSVT)
The challenge focuses on exploring the limits of deep learning models and algorithms under the situations of large-scale data, and advances innovation in large-scale scene text recognition in partial annotations.
Derived from Baidu's real-world scenarios and as the first ever scene text dataset labeled with partial annotations, ICDAR2019-LSVTconsists of 20,000 testing data, 30,000 training data in full annotations and 400,000 training data in partial annotations, at least 14 times as large as existing robust reading benchmarks (ICDAR 2017, ICPR 2018, etc.).
Scene text reading in natural images has a wide range of applications including camera translation, image retrieval, street view and landmark recognition, outdoor scene understanding and more.
50,000 street view images in full annotations and 400,000 street view images in partial annotationof a large variety of complicated real-world scenarios.
Text detection and end-to-end text recognition both awarding a prize of $8,700
$8,700 in total, sponsored by Baidu.
Task 1. Text detection, $2,500/$1,250/$600 for top 3 winners.
Task 2. End-to-end text spotting, $2,500/$1,250/$600 for top 3 winners.
ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text (ICDAR 2019-ArT)
The ICDAR2019-ArTdataset is composed of Total-Text, SCUT-CTW1500, and Baidu Curved Scene Text There is a total of 10,176 images, which makes ArTone of the larger scale scene text datasets today and an unique dataset since most of the existing datasets were dominated by horizontal and multi-oriented text instances only.
The scene text detection, recognition, and spotting problems of arbitrarily-shaped text poses more of a challenge for existing algorithms than regular text recognition. Therefore, ICDAR 2019-ArT aims to fuel the innovation of ways to improve the scene text understating of arbitrarily-shaped text.
· Data set
10,000 labels with arbitrarily-shaped text in both Chinese in English; the largest data set of arbitrarily-shaped text
$8,700 in total, sponsored by Baidu.
Task 1. Scene Text Detection, $1,700/$800/$400 for top 3 winners.
Task 2. Scene Text Recognition, $1,700/$800/$400 for top 3 winners.
Task 3. Scene Text Spotting, $1,700/$800/$400 for top 3 winners.
University of Malaya
South China University of Technology
Computer Vision Centre, Spain
China Society of Image & Graphics
February 21 to March 31, 2019: Registration open
March 1, 2019: Training module open for download
April 9, 2019: Submission starting time
April 30, 2019: Submission closing time
ICDAR2019 Robust Reading Challenge on Large-scale Street View Text with Partial Labeling:
ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text: