GraphKD (Best Student Paper ICDAR, 2024,🏆)

Description

Pytorch implementation of the paper Exploring Knowledge Distillation Towards Document Object Detection with Structured Graph Creation. This model is implemented on top of the detectron2 framework. The proposed architecture explores graph-based knowledge distillation to mitigate the trade-off between no. of model parameters (trainable) and performance accuracy towards document knowledge distillation with adaptive node sampling strategy and weighted edge distillation via Mahalanobis distance.

Structured graph creation: We extracted the RoI pooled features and classified them into "Text" and "Non-text" based on their covariance. Then we initialize the node in the identified RoI regions and define the adjacency edges. Lastly, we iteratively merge the text node with an adaptive sample mining strategy to reduce text bias.

Getting Started

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

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

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

conda create --name graphkd python=3.9
conda activate graphkd
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git' --user

Step 3: Running

for training:

./start_train.sh train projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

for testing:

./start_train.sh eval projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

for debugging:

./start_train.sh debugtrain projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

Results and Model Zoo

1. PublayNet

Model	Config-file	Weights	AP
R50-R18	config-publay	model	28.0
R101-R50	config-publay	model	88.6
R152-R101	config-publay	model	88.8
R101-EB0	config-publay	model	27.6
R50-MNv2	config-publay	model	28.2

2. Prima

Model	Config-file	Weights	AP
R50-R18	config-prima	model	26.5
R101-R50	config-prima	model	35.0
R152-R101	config-prima	model	41.9
R101-EB0	config-prima	model	12.6
R50-MNv2	config-prima	model	14.9

3. Historical Japanese

Model	Config-file	Weights	AP
R50-R18	config-prima	model	33.4
R101-R50	config-prima	model	78.3
R152-R101	config-prima	model	79.7
R101-EB0	config-prima	model	33.1
R50-MNv2	config-prima	model	37.5

4. DoclayNet

Model	Config-file	Weights	AP
R50-R18	config-prima	model	42.1
R101-R50	config-prima	model	65.0
R152-R101	config-prima	model	68.9
R101-EB0	config-prima	model	28.9
R50-MNv2	config-prima	model	23.6

Citation

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

@article{banerjee2024graphkd,
  title={GraphKD: Exploring Knowledge Distillation Towards Document Object Detection with Structured Graph Creation},
  author={Banerjee, Ayan and Biswas, Sanket and Llad{\'o}s, Josep and Pal, Umapada},
  journal={arXiv preprint arXiv:2402.11401},
  year={2024}
}

Acknowledgement

We have built it on the top of the Dsig.

Conclusion

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