Table Detection Modules

A collection of table detection modules designed to accurately detect tables across various document types, such as invoices, scientific papers, and other complex layouts.

Overview

This repository provides multiple table detection models to efficiently locate and identify tables within documents of diverse formats. We aim to continuously update and improve each module. Our goal is to establish a cohesive framework that allows all models to be seamlessly called from a unified script.

Models Implemented and Tested

Our modules currently include the following table detection models:

1. YOLO - You Only Look Once, a real-time object detection system
2. DETR - DEtection TRansformers for high-quality table detection
3. CascadeTabNet - Cascade Mask R-CNN-based model optimized for table structure recognition

Training Dataset

Our models are primarily trained on Doclaynet, a comprehensive human-annotated document layout segmentation dataset with 80,863 pages from a wide variety of sources. Additional testing has been conducted on complex proprietary medical documents to validate robustness and versatility across various domains.

Model Performance

Model	Train Dataset	Test Dataset	mAP50	mAP50:95
DETR	DoclayNet	DoclayNet	0.94	0.83
LayoutLMv3	DoclayNet	DoclayNet	0.91	0.86
CascadeTabNet	DoclayNet	DoclayNet	0.90	0.82
YOLO	DoclayNet	DoclayNet	0.92	0.88

The below mentioned results are from the model which were first trained on Doclaynet and then fine tuned for proprietary data.

Model	Train Dataset	Test Dataset	mAP50	mAP50:95
DETR	Proprietary	Proprietary	0.82	0.57
LayoutLMv3	Proprietary	Proprietary	0.91	0.74
CascadeTabNet	Proprietary	Proprietary	0.76	0.52
YOLO	Proprietary	Proprietary	0.90	0.83
YOLO	Proprietary	Proprietary 2	0.93	0.83

Acknowledgments

Special thanks to the following resources and frameworks that have significantly supported our development:

1. YOLO Documentation
2. DETR by Facebook AI Research
3. MMdetection Toolkit

Contact

For questions, feedback, or collaboration opportunities, please feel free to reach out:

• Anik De
• Ritu Singh