Awesome

Data Downloading and Processing Pipeline for OVEN

• Community Contributed Repository
• [Project Page][Paper] [Official Dataset Page][Sibling Project (Infoseek)]

Open-domain Visual Entity Recognition: Towards Recognizing Millions of Wikipedia Entities (ICCV 2023 Oral)

Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, Ming-Wei Chang.

Release

• [24/5/30] Image data is on Huggingface dataset now 🤗: ychenNLP/oven.
• [23/6/10] Implementing the data downloading script that can reproduce the data from original images.
• [23/6/7] Releasing the annotation for the OVEN Dataset.

OVEN Dataset

• [24/9/1] We have switched data download from Google Drive to Huggingface datasets due to storage limitation issue.

To download annotations and an image snapshot, check out Huggingface dataset at 🤗: ychenNLP/oven, which support high-speed wget command.

• To download Wikipedia 6M knowledge base (title only or image url), please run bash script download_wiki.sh. Download 6M wiki infobox images from 'all_wikipedia_images.tar'.

To download all images from the source dataset, please go to "image_downloads/" and run all download scripts. Then run the following script to merge all data with ovenid2impath.csv:

python merge_oven_images.py

Evaluation

python run_oven_eval.py

# ===== BLIP2 Zeroshot ====
# ===== Validation ========
# ===== Final score 7.87
# ===== Query Split score 20.58
# ===== Entity Split score 4.87
# ===== Query Seen Accuracy 24.63
# ===== Query Unseen Accuracy 17.68
# ===== Entity Seen Accuracy 8.55
# ===== Entity Unseen Accuracy 3.4

Starting Code

• Run BLIP2 zero-shot inference:

python run_blip2_oven.py --split val_entity

• Next, we need to run BM25 to map the BLIP2 predictions to Wikipedia 6M label space:

python run_bm25_query.py --input_file {INPUT} --output_file {OUTPUT}

• Before running BM25, you need to run BM25 index of Wikipedia (Download Wikipedia from the "Wiki6M_ver_1_0_title_only.jsonl")

python run_bm25_index.py

Acknowledgement

If you find OVEN useful for your research and applications, please cite using this BibTeX:

@article{hu2023open,
  title={Open-domain Visual Entity Recognition: Towards Recognizing Millions of Wikipedia Entities},
  author={Hu, Hexiang and Luan, Yi and Chen, Yang and Khandelwal, Urvashi and Joshi, Mandar and Lee, Kenton and Toutanova, Kristina and Chang, Ming-Wei},
  journal={arXiv preprint arXiv:2302.11154},
  year={2023}
}