DocSegTr

Description

Official Pytorch implementation of the paper DocSegTr: An Instance-Level End-to-End Document Image Segmentation Transformer. This model is implemented on top of the adelaidet and detectron2 frameworks. The paper proposes a novel bottom-up instance segmentation strategy using Transformers to segment instances(document layouts) in scientific document images from the PubLayNet benchmark.

DocSegtr builds on a simple CNN feature extractor with FPN on the input document image. The multi-scaled feature maps(P2-P6) from FPN are combined with positional embedding information to feed into transformer layers, to predict document instances and generate corresponding kernel dynamically. The layerwise feature aggregation module combines the local FPN features and global transformer feature from P5 to segment the instances on the document image

Getting Started

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

git clone https://github.com/biswassanket/DocSegTr.git 
cd DocSegTr

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

conda env create -f environment.yml
conda activate instaseg

Step 3: Build detectron 2 and adelaidet from source:

Building detectron2 v0.2.1 from source using the following link to download.

cd detectron2-0.2.1
python setup.py build develop

To build adelaidet from source you need this simple command:

cd .. //going back to original work dir
python setup.py build develop

Step 4: Downloading dataset

• To download PubLayNet dataset: curl -o <YOUR_TARGET_DIR>/publaynet.tar.gz https://dax-cdn.cdn.appdomain.cloud/dax-publaynet/1.0.0/publaynet.tar.gz

Step 5: For testing our model, download the best pretrained model weights from this link best model

python tools/train_net_custom.py \
    --config-file configs/SOTR/R_101_DCN_doc.yaml \
    --eval-only \
    --num-gpus 1 \
    MODEL.WEIGHTS work_dir/.../model_final.pth

Step 6: For training the model from scratch, use this magic command for training on 'n' GPUs:

python tools/train_net_custom.py \
    --config-file configs/SOTR/R_101_DCN_doc.yaml \
    --num-gpus 2

Step 7: For visualizing qualitative result predictions on PubLayNet, use the following:

python tools/visualize_publaynet.py \
    --input /path to JSON created by trained model/ \
    --output /path to output_dir \
    --dataset publaynet_minival \
    --conf-threshold 0.6

Results

Qualitative analysis on the PubLayNet dataset by DocSegTr . Here first, second and third columns represent original image, ground truth and our proposed DocSegTr results, respectively.

Citation

If you find this code useful in your research then please cite

@article{biswas2022docsegtr,
  title={DocSegTr: An Instance-Level End-to-End Document Image Segmentation Transformer},
  author={Biswas, Sanket and Banerjee, Ayan and Llad{\'o}s, Josep and Pal, Umapada},
  journal={arXiv preprint arXiv:2201.11438},
  year={2022}
}

Acknowledgement

Our project has adapted and borrowed the code structure from SOTR. We thank the authors. This research has been partially supported by the Spanish projects RTI2018-095645-B-C21, and FCT-19-15244, and the Catalan projects 2017-SGR-1783, the CERCA Program / Generalitat de Catalunya and PhD Scholarship from AGAUR (2021FIB-10010).

Author

• Sanket Biswas
• Ayan Banerjee

Conclusion

Thank you and sorry for the bugs!