Awesome
NewsStories: Illustrating articles with visual summaries
This repository contains a PyTorch implementation of the paper NewsStories: Illustrating articles with visual summaries accepted at ECCV 2022. If you find this implementation or the paper helpful, please consider citing:
@InProceedings{rxtan2022newsstories,
author={Tan, Reuben and Plummer, Bryan and Saenko, Kate and Lewis, JP and Sud, Avneesh and Leung, Thomas},
title={NewsStories: Illustrating articles with visual summaries},
booktitle={Proceedings of the IEEE/CVF European Conference on Computer Vision (ECCV)},
year={2022} }
Dependencies
- Python 3.6
- Pytorch version 1.10.0
- nltk
Download NewStories Dataset
Please use this here to download the full dataset which contains the meta data required for downloaading the images, articles and videos. The evaluation split with 5 images can also be found here.
Download GoodNews Dataset and evaluation splits
To begin, please follow the instructions here to download the GoodNews dataset. In our experiments, we evaluate the trained models on the GoodNews dataset.
Next, you can download the evaluation splits that we use in our experiments using the links below:
MIL-SIM code
Training code on NewsStories
python train_model.py --pretrained_path {path to pretrained CLIP model} --num_imgs {maximum number of images in a set} --num_sentences {maximum number of sentences in an article.}
Evaluation code on GoodNews
python eval_goodnews.py
Required Arguments For Training
- pretrained_path: path to pretrained CLIP model
- num_imgs: maximum number of images in a set
- num_sentences: maximum number of sentences in an article.
Contact
Please do not hesitate to contact me at rxtan@bu.edu if you have any questions or problems running the code.