TableBank

We have trained several models for the LaTeX part of the TableBank dataset using the Mask-RCNN, RetinaNet, SSD and YOLO algorithms using different deep learning libraries. To this aim, we have used several libraries:

• Keras library for Mask RCNN
• Keras library for RetinaNet
• Implementation of SSD in MXNet
• Implementation of YOLO in Darknet

Results

To evaluate our models, we have employed the same metric used in the ICDAR 2019 Competition on Table Detection.

Model	P@0.6	R@0.6	F1@0.6	P@0.7	R@0.7	F1@0.7	P@0.8	R@0.8	F1@0.8	P@0.9	R@0.9	F1@0.9	WAvgF1
Mask RCNN	0,94	0,98	0,96	0,94	0,97	0,95	0,93	0,96	0,94	0,84	0,87	0,86	0,92
RetinaNet	0,98	0,86	0,92	0,98	0,86	0,92	0,97	0,85	0,91	0,94	0,82	0,87	0,90
SSD	0,96	0,97	0,96	0,94	0,95	0,95	0,92	0,92	0,92	0,82	0,82	0,82	0,90
YOLO	0,98	0,99	0,98	0,98	0,99	0,98	0,96	0,97	0,96	0,74	0,75	0,75	0,90

Model Zoo

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

• Mask RCNN: weights.
• RetinaNet: weights, classes file.
• SSD: weights.
• YOLO: weights, config file, names file.

Colab Notebooks for prediction

You can use the trained models with the following notebooks.

• MaskRCNN Notebook
• RetinaNet Notebook
• SSD Notebook
• YOLO Notebook