(GVIL) - Visual Illusion Dataset

• Check out project homepage
• Check out the full paper

Setup

1. Unzip the data.

unzip dataset.zip

PS: The dataset is also available on Huggingface: link.

2. Install some utilities.

pip install matplotlib tqdm

Inference

1. Add your model in model.py.
2. Add the build function in inference.py.
3. Run inference.

python inference.py --task vqa --model YOUR_MODEL_NAME

Evaluation

We provide the model predictions from OFA and Unified-IO for reference. For example, you can run evaluation on the prediction of the OFA-Large model after unzipping unzip predictions.zip.

python eval.py \
    --vqa_predictions predictions/vqa__ofa_large.json \
    --vg_predictions predictions/vg__ofa_large.json

Citation

@inproceedings{zhang2023grounding,
    title={Grounding Visual Illusions in Language: Do Vision-Language Models Perceive Illusions Like Humans?},
    author={Zhang, Yichi and Pan, Jiayi and Zhou, Yuchen and Pan, Rui and Chai, Joyce},
    booktitle={Proceedings of Conference of Empirical Methods in Natural Language Processing},
    year={2023},
    organization={EMNLP 2023}
}