Related Experiment Video
Updated: Aug 9, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.3K
A large-scale dataset for end-to-end table recognition in the wild
Fan Yang1, Lei Hu1, Xinwu Liu2
1School of Electronic and Information Engineering, South China University of Technology, Guangzhou, 510641, China.
Scientific Data
|February 24, 2023
Summary
Researchers introduce TabRecSet, a large-scale, bilingual dataset for end-to-end table recognition (TR). This dataset addresses the lack of benchmarks, enabling simultaneous table detection, structure recognition, and content recognition in diverse real-world scenarios.
Area of Science:
- Computer Vision
- Pattern Recognition
- Machine Learning
Background:
- Table recognition (TR) involves table detection (TD), table structure recognition (TSR), and table content recognition (TCR).
- End-to-end TR, performing all sub-tasks simultaneously in real-world scenarios, remains an unexplored research area due to the absence of a suitable benchmark dataset.
- Existing datasets often lack comprehensive annotations or struggle with diverse table forms and real-world image conditions.
Purpose of the Study:
- To introduce a novel, large-scale, bilingual dataset, TabRecSet, specifically designed for end-to-end table recognition research.
- To provide a comprehensive benchmark that facilitates simultaneous TD, TSR, and TCR for diverse and irregular tables found in various real-world scenarios.
- To develop an efficient and high-quality annotation tool, TableMe, to support dataset creation.
Main Methods:
- Creation of TabRecSet, a large-scale dataset featuring 38.1K tables (20.4K English, 17.7K Chinese) with diverse forms (e.g., incomplete, irregular, rotated) and scenarios (scanned, camera-taken, documents, invoices).
- Comprehensive annotation including spatial (polygon-based) and logical cell annotations, and text content for TD, TSR, and TCR.
- Development of TableMe, a visualized and interactive annotation tool to enhance annotation efficiency and quality.
Main Results:
- TabRecSet is the largest and first bilingual dataset specifically curated for end-to-end table recognition.
- The dataset includes polygon-based spatial annotations, which are more effective for irregular tables compared to traditional bounding boxes or quadrilaterals.
- The TableMe tool demonstrated improved efficiency and quality in the annotation process.
Conclusions:
- The introduction of TabRecSet significantly advances end-to-end table recognition research by providing a much-needed benchmark.
- The dataset's diversity in table forms and scenarios, coupled with polygon annotations, prepares models for real-world complexities.
- TabRecSet and the TableMe tool collectively pave the way for more robust and accurate table information extraction systems.
