Publaynet

• Publaynet: Largest Dataset Ever for Document Layout Analysis

Company Articles DataSet

Overview

Category	Training Set	Validating Set	Testing Set
Num of Images	20365	500	499
Percentage	95%	2.5%	2.5%

Training Set:

category	#instances	category	#instances	category	#instances	category	#instances
chapter	11312	section	17471	clause	106931	total	135714

Validating Set:

category	#instances	category	#instances	category	#instances	category	#instances
chapter	151	section	246	clause	3096	total	3493

Testing Set:

category	#instances	category	#instances	category	#instances	category	#instances
chapter	151	section	249	clause	2947	total	3347

Download

All Files:

Images

• 001-100
• 101-200
• 201-300
• 301-400
• 401-500

Annotation

• Txt
• Json

Dataset:

• Beihang Pan:

  • Training Set
  • Validating Set
  • Testing Set

• Google Drive:

  • Training Set
  • Validating Set
  • Testing Set

Model:

Pretrained on Publaynet Dataset

• Model finetuned with Publaynet Dataset based on pretrained model of Faster-RCNN-ResNet

Trained on Company Articles Dataset

• Best Model finetuned with Company Articles Dataseton based on pretrained model of Faster-RCNN-ResNet

Python Files:

• faster_rcnn_resnet101_coco_2018_01_28: backbone的预训练模型，用于publaynet数据集训练
• visualizeSet.py: 可视化数据集
• build.py: 构建优化器和学习率策略
• utils.py: 使用publaynet数据集的工具文件
• train.py: 使用publaynet数据集的训练文件
• test_per_img.py: 可视化测试集的预测结果
• predict.py: 使用publaynet数据集的预测文件

Requirements

Detectron2

• Linux or macOS with Python ≥ 3.6
• cython
• opencv-python
• torchvision　(PyTorch ≥ 1.3)
• 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

Run on Google Colab:

Install Requirements and Clone Publaynet

!pip install pyyaml==5.1
!pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
!git clone https://github.com/noba1anc3/Publaynet.git
cd Publaynet

Build Detectron2 from Source

After having the above dependencies and gcc & g++ ≥ 5, run:

!git clone https://github.com/facebookresearch/detectron2.git
cd detectron2
!python -m pip install -e .
cd ..

# Or if you are on macOS
# CC=clang CXX=clang++ python -m pip install -e .

Train

Data Preparation

Mount Google Drive

from google.colab import drive
drive.mount('/content/drive/')

Copy Training and Testing Data to Publaynet's Path

mkdir data

cp -rf ../drive/'My Drive'/train.zip ./data/
cp -rf ../drive/'My Drive'/val.zip ./data/

cd data
!unzip train.zip
!unzip val.zip
cd ..

Finetune on Faster_RCNN_X_101_32x8d_FPN_3x

!python train.py -f False

Finetune on Publaynet's Pretrained Model

mkdir output
cp -rf ../drive/'My Drive'/model_final.pth ./output/
!python train.py -f True

Training Log

• Evaluation Norm

Training From Scratch

• Training Log

Training on Faster-RCNN Pretrained Model

• Training Log
• Loss Json File
• TensorBoard Log File
• Best Model

Training on Pretrained Model Finetuned on Publaynet Dataset

• Training Log

Comparison

Training From Scratch & Training on Faster RCNN Pretrained Model

Faster RCNN Pretrained Model & Publaynet Pretrained Model

Evaluation Result on Testing Set

Per-class AP

chapter AP	section AP	clause AP	mAP
85.180	86.641	93.367	88.396

Average Precision

AP	AP50	AP75	APs	APm	APl
88.396	99.037	98.956	NaN	80.382	88.964

Average Recall

AR1	AR10	AR100	ARs	ARm	ARl
57.0	91.4	92.0	NaN	84.8	92.1