FoodieQA: A Multimodal Dataset for Fine-Grained Understanding of Chinese Food Culture

Github Repo

😋 We release all tools and code used to create the dataset at https://github.com/lyan62/FoodieQA.

Paper

For more details about the dataset, please refer to 📄 FoodieQA: A Multimodal Dataset for Fine-Grained Understanding of Chinese Food Culture

Dataset Download

!!Note!! The Json files are in the FoodieQA.zip (click on the Files and Versions tab to download), or download the dataset directly with git clone.

Terms and Conditions for Data Usage

By downloading and using the data, you acknowledge that you have read, understood, and agreed to the following terms and conditions.

1. Research Purpose: The data is provided solely for research purposes and must not be used for any commercial activities.

2. Evaluation Only: The data may only be used for evaluation purposes and not for training models or systems.

3. Compliance: Users must comply with all applicable laws and regulations when using the data.

4. Attribution: Proper attribution must be given in any publications or presentations resulting from the use of this data.

5. License: The data is released under the CC BY-NC-ND 4.0 license. Users must adhere to the terms of this license.

Data Structure

• /images: contains all images needed for multi-image VQA and single-image VQA task.

• mivqa_tidy.json questions for Multi-image VQA task.

  • data format

    {
        "question": "哪一道菜适合喜欢吃肠的人？",
        "choices": "",
        "answer": "0",
        "question_type": "ingredients",
        "question_id": qid,
        "ann_group": "闽",
        "images": [
            img1_path, img2_path, img3_path, img4_path
        ],
        "question_en": "Which dish is for people who like intestine?"
    }
    

• sivqa_tidy.json question for Single-image VQA task.

  • data format

    {
        "question": "图片中的食物是哪个地区的特色美食?",
        "choices": [
            ...
        ],
        "answer": "3",
        "question_type": "region-2",
        "food_name": "梅菜扣肉",
        "question_id": "vqa-34",
        "food_meta": {
            "main_ingredient": [
                "肉"
            ],
            "id": 253,
            "food_name": "梅菜扣肉",
            "food_type": "客家菜",
            "food_location": "餐馆",
            "food_file": img_path
        },
        "question_en": translated_question,
        "choices_en": [
            translated_choices1,
            ...
        ]
    }
    

• textqa_tidy.json

  • data format

    {
        "question": "酒酿圆子属于哪个菜系?",
        "choices": [
            ...
        ],
        "answer": "1",
        "question_type": "cuisine_type",
        "food_name": "酒酿圆子",
        "cuisine_type": "苏菜",
        "question_id": "textqa-101"
    },
    

Models and results for the VQA tasks

Evaluation	Multi-image VQA (ZH)	Multi-image VQA (EN)	Single-image VQA (ZH)	Single-image VQA (EN)
Human	91.69	77.22†	74.41	46.53†
Phi-3-vision-4.2B	29.03	33.75	42.58	44.53
Idefics2-8B	50.87	41.69	46.87	52.73
Mantis-8B	46.65	43.67	41.80	47.66
Qwen-VL-12B	32.26	27.54	48.83	42.97
Yi-VL-6B	-	-	49.61	41.41
Yi-VL-34B	-	-	52.73	48.05
GPT-4V	78.92	69.23	63.67	60.16
GPT-4o	86.35	80.64	72.66	67.97

Models and results for the TextQA task

Model	Best Accuracy	Prompt
Phi-3-medium	41.28	1
Mistral-7B-instruct	35.18	1
Llama3-8B-Chinese	47.38	1
YI-6B	25.53	3
YI-34B	46.38	3
Qwen2-7B-instruct	68.23	3
GPT-4	60.99	1

BibTeX Citation

@article{li2024foodieqa,
  title={FoodieQA: A Multimodal Dataset for Fine-Grained Understanding of Chinese Food Culture},
  author={Li, Wenyan and Zhang, Xinyu and Li, Jiaang and Peng, Qiwei and Tang, Raphael and Zhou, Li and Zhang, Weijia and Hu, Guimin and Yuan, Yifei and S{\o}gaard, Anders and others},
  journal={arXiv preprint arXiv:2406.11030},
  year={2024}
}