Further information and evaluations can be found in the paper
We tested our code on a Linux machine with python 3.10, detectron 0.6 and pytorch 1.11.
Installation instructions for detectron and pytorch can be found here
To setup the environment with all the required dependencies, we provide further steps here
Please use following link to download model checkpoints and datasets.
Unzip checkpoints.zip and datasets.zip at the root level of this repository and download the images as described in download_ep_images_helper. Move the train/test/val image directories to datasets/eperiodica3/imgs.
At the moment, there are two images which are inaccessible to the public due to copyright restrictions. Until they are publicly available, we download similar images from these magazines for which the original bounding boxes roughly match. In 2024 "edm.001.2018.073.0201-0" in the training set will be publicly available, and "tbg.002.2020.158.0072-0" in the test set will be publicly available in 2026.
DSG_E2E_arxivdocs: DSG trained on arXivdocs
DSG_E2E_eperiodica: DSG trained on E-Periodica
Note: When running the code for the first time, glove word embeddings are automatically downloaded.
First, create an output directory, e.g. at ./demo/EP_outputs.
To run DSG for prediction and use grammar-based postprocessing, run:
python visualizations/demo.py --config-file ./configs/sgg_end2end_EP.yaml --input ./datasets/eperiodica3/imgs/val/* --output ./demo/EP_outputs --raw_output ./demo/EP_outputs --opts MODEL.ROI_SCENEGRAPH_HEAD.PREDICT_USE_VISION True MODEL.WEIGHTS ./checkpoints/DSG_E2E_eperiodica/dsg_e2e_eperiodica_checkpoint.pth TEST.USE_GRAMMAR_POSTPROCESSING True
The hOCR creation demo uses the outputs created by the previous script.
For convenience, we prepared outputs for one sample and a jupyter notebook to demonstrate our hOCR creation and querying here
This repository builds on other open source implementations, including detectron2 and segmentation-sg