Model Zoo for table detection

We have trained several models for the following datasets: ICDAR13, ICDAR17 (this dataset is used to construct models for detecting tables, figures and formulas), ICDAR19 modern images, Invoices (a private dataset that is available under request), Marmot Chinese, Marmot English and UNLV. Since these datasets do not provide a publicly available test set; we have split the training sets using the partition 75% for training and 25% for testing. The dataset splits are available as follows.

• ICDAR13: train set, test set.
• ICDAR17: train set, test set.
• ICDAR17FIG: train set, test set.
• ICDAR17FOR: train set, test set.
• ICDAR19: train set, test set.
• Invoices: train set, test set.
• MarmotChi: train set, test set.
• MarmotEn: train set, test set.
• UNLV: train set, test set.

Tablebank weights

• Mask-RCNN weights.
• RetinaNet weights.
• SSD weights.
• YOLO weights.

You can see how to fine-tune models using the above weights in the fine-tuning page.

Fine-tuning from natural images

From models trained with the Pascal VOC dataset, we have created several detection models for the aforementioned datasets using fine-tuning with the algorithms Mask-RCNN, RetinaNet, SSD and YOLO. The results are summarized in the following figure.

The trained models are available in the format used by each framework, and distributed under the GNU General Public License v3.0

	Mask RCNN	RetinaNet	SSD	YOLO
ICDAR13	weights	weights,classes file	weights	weights, config file, names file
ICDAR17	weights	weights,classes file	weights	weights, config file, names file
ICDAR17FIG	weights	weights,classes file	weights	weights, config file, names file
ICDAR17FOR	weights	weights,classes file	weights	weights, config file, names file
ICDAR19	weights	weights,classes file	weights	weights, config file, names file
Invoices	weights	weights,classes file	weights	weights, config file, names file
MarmotChi	weights	weights,classes file	weights	weights, config file, names file
MarmotEn	weights	weights,classes file	weights	weights, config file, names file
UNLV	weights	weights,classes file	weights	weights, config file, names file

Fine-tuning from the TableBank dataset

From models trained with the TableBank dataset, we have created several detection models for the aforementioned datasets using fine-tuning with the algorithms Mask-RCNN, RetinaNet, SSD and YOLO. The results are summarized in the following figure.

The trained models are available in the format used by each framework, and distributed under the GNU General Public License v3.0

	Mask RCNN	RetinaNet	SSD	YOLO
ICDAR13	weights	weights,classes file	weights	weights, config file, names file
ICDAR17	weights	weights,classes file	weights	weights, config file, names file
ICDAR17FIG	weights	weights,classes file	weights	weights, config file, names file
ICDAR17FOR	weights	weights,classes file	weights	weights, config file, names file
ICDAR19	weights	weights,classes file	weights	weights, config file, names file
Invoices	weights	weights,classes file	weights	weights, config file, names file
MarmotChi	weights	weights,classes file	weights	weights, config file, names file
MarmotEn	weights	weights,classes file	weights	weights, config file, names file
UNLV	weights	weights,classes file	weights	weights, config file, names file

Comparison of models

In the next table, we compare the WF1Avg score obtained by the models fine-tuned from models constructed using natural images, and the models constructed using the TableBank dataset.

	Mask R-CNN Natural	Mask R-CNN TableBank	RetinaNet Natural	RetinaNet TableBank	SSD Natural	SSD TableBank	YOLO Natural	YOLO TableBank
ICDAR13	0.09	0.72	0.5	0.74	0.4	0.52	0.50	0.66
ICDAR17	0.27	0.72	0.66	0.84	0.45	0.47	0.72	0.82
ICDAR17FIG	0.19	0.29	0.72	0.72	0.59	0.29	0.63	0.71
ICDAR17FOR	0.09	0.06	0.1	0.2	0.32	0.35	0.52	0.59
ICDAR19	0.31	0.65	0.64	0.72	0.2	0.23	0.81	0.85
Invoices	0.22	0.44	0.65	0.66	0.59	0.66	0.63	0.69
MarmotChi	0.21	0.70	0.69	0.84	0.58	0.64	0.7	0.87
MarmotEn	0.39	0.82	0.70	0.8	0.44	0.45	0.83	0.85
UNLV	0.15	0.55	0.73	0.74	0.42	0.49	0.7	0.77