Description

Pytorch implementation of the paper SwinDocSegmenter: An End-to-End Unified Domain Adaptive Transformer for Document Instance Segmentation. This model is implemented on top of the detectron2 framework. The proposed model can be used to analyze complex layouts including magazines, Scientific Reports, historical documents, patents and so on as shown in the following examples.

Magazines	Scientific Reports
Tables	Others

Getting Started

Step 1: Clone this repository and change directory to repository root

git clone https://github.com/ayanban011/SwinDocSegmenter.git 
cd SwinDocSegmenter

Step 2: Setup and activate the conda environment with required dependencies:

follow the installation instructions

Step 3: For testing our model, download the best pretrained model weights from the Model Zoo

python ./train_net.py \
    --config-file maskdino_R50_bs16_50ep_4s_dowsample1_2048.yaml \
    --eval-only \
    --num-gpus 1 \
    MODEL.WEIGHTS ./model_final.pth

Step 4: For training the model from scratch, use this magic command for training on 'n' GPUs:

python train_net.py --num-gpus 1 --config-file config_path SOLVER.IMS_PER_BATCH SET_TO_SOME_REASONABLE_VALUE SOLVER.BASE_LR SET_TO_SOME_REASONABLE_VALUE

Step 4: For training the model from scratch, use this magic command for training on 'n' GPUs:

In train_net.py

def main(args):
    register_coco_instances("dataset_train",{},"path to the ground truth json file","path to the training image folder")
    register_coco_instances("dataset_val",{},"path to the ground truth json file","path to the validation image folder")

    MetadataCatalog.get("dataset_train").thing_classes = ['name of the classes']
    MetadataCatalog.get("dataset_val").thing_classes = ['name of the classes']
    ...
if __name__ == "__main__":
    ...
    MetadataCatalog.get("dataset_train").thing_classes = ['name of the classes']
    MetadataCatalog.get("dataset_val").thing_classes = ['name of the classes']
    ...

In Config File

...
SEM_SEG_HEAD:
    ...
    NUM_CLASSES: #no. of classes
...
DATASETS:
  TRAIN: ("dataset_train",)
  TEST: ("dataset_val",)
...

Model Zoo

In this section, we release the pre-trained weights for all the best SwinDocSegmenter model variants trained on benchmark datasets.

Dataset	Config-file	Weights	AP
PublayNet	config-publay	model	93.72
Prima	config-prima	model	54.39
HJ Dataset	config-hj	model	84.65
TableBank	config-table	model	98.04
DoclayNet	config-doclay	model	76.85

Test on Custom Images

In order to, test the custom images, you should put all the images and create a JSON file that contains all the categories of the dataset; image_name; image_id. We have created a sample test set for DocLayNet. You can create the same for other datasets, by following the same structure.

Dataset Link: Custom DocLayNet Test Set

Annotations Link: A little annotations

For Semi-Supervised Settings

Please refer to the SemiDocSeg folder for generating support sets and new_annotations. The rest of the training and testing procedure is as same as SwinDocSegmenter.

Citation

If you find this useful for your research, please cite it as follows:

@article{banerjee2023swindocsegmenter,
  title={SwinDocSegmenter: An End-to-End Unified Domain Adaptive Transformer for Document Instance Segmentation},
  author={Banerjee, Ayan and Biswas, Sanket and Llad{\'o}s, Josep and Pal, Umapada},
  journal={arXiv preprint arXiv:2305.04609},
  year={2023}
}

@article{banerjee2024semidocseg,
  title={SemiDocSeg: harnessing semi-supervised learning for document layout analysis},
  author={Banerjee, Ayan and Biswas, Sanket and Llad{\'o}s, Josep and Pal, Umapada},
  journal={International Journal on Document Analysis and Recognition (IJDAR)},
  pages={1--18},
  year={2024},
  publisher={Springer}
}

Acknowledgement

Many thanks to these excellent opensource projects

• MaskDINO
• DINO

Authors

• Ayan Banerjee
• Sanket Biswas

Conclusion

Thank you for your interest in our work, and sorry if there are any bugs.