README.md

---
license: other
task_categories:
- image-classification
- object-detection
language:
- en
tags:
- food
- meal
- multi-object
- bounding-box
size_categories:
- 1K<n<10K
extra_gated_heading: "Academic-Only Dataset Access Request"
extra_gated_description: "This dataset is available exclusively to academic organizations for non-commercial use. Please answer the following questions to verify your eligibility."
extra_gated_prompt: "Eligibility is limited to individuals affiliated with academic institutions using this dataset for research or educational purposes only."

extra_gated_fields:
  Are you affiliated with an academic or research institution?:
    type: select
    options:
      - Yes
      - No
  Is your intended use of the dataset strictly non-commercial (e.g., research or education)?:
    type: select
    options:
      - Yes
      - No
  Name of your institution or university: text
  Official email address (with institutional domain): text
  Brief description of intended use: text
---
# FoodLogAthl Meal Image Dataset

## Overview  
FoodLogAthl-218 is a **real-world meal image** dataset containing **6,925 original photos**, each annotated with **one or more food bounding boxes** and rich **metadata** (capture date, user ID). Unlike web-sourced collections, these images come from an actual dietary management app, FoodLog Athl, reflecting everyday photo conditions and meal diversity.

## Background & Motivation  
Most public food image datasets (e.g. Food-101, UEC Food256, Food2K) are built from carefully curated web images, often favoring visually polished, single-dish photos. Such benchmarks fail to capture the messy, multi-dish snapshots people take for personal food logs.  
FoodLogAthl-218 bridges this gap by leveraging images uploaded by real users of the FoodLog Athl app, yielding a challenging, authentic benchmark for both classification and object-detection models in dietary-management applications.

## Data Collection  
- **Source**: FoodLog Athl (dietary-management app) 
- **Period**: May 2023 – Oct 2024
- **User selection**: all meal records linked to dietitians with ≥3 active users  
- **Record contents**: user ID, recording date, original image, crop-based bounding boxes, dish names
- **Multi-dish handling**: each detected dish becomes a separate sample, but images with any invalid crop are discarded as a whole to preserve real meal frequency

## Directory Structure  
```
FoodLogAthl-218/
├── images/
│   ├── image_XXXXX.jpg
│   └── ... (total 6,925 images)
├── annotations/
│   └── instances_coco.json
├── metadata.csv
├── class_map.csv
├── .gitattributes
├── .gitignore
└── README.md
```

## File Descriptions  
- **images/**  
  All original `.jpg` files named `image_<ID>.jpg`.  
- **annotations/instances_coco.json**  
  Standard COCO fields:  
  - `images` (id, file_name, width, height)  
  - `annotations` (id, image_id, category_id, bbox [x_min, y_min, w, h], area, iscrowd)  
  - `categories` (id, name)  
- **metadata.csv**  
  | file_name        | record_date | user_id |  
  |------------------|-------------|---------|  
  | image_125683.jpg | 2024-06-17  | 248     |  
  | image_121223.jpg | 2023-05-12  | 105     |  
  | image_123574.jpg | 2024-06-17  | 105     |   
  | …                | …           | …       |  
- **class_map.csv**: 
  - Mapping information between class names and IDs
- **.gitattributes**: 
  - Git attributes configuration file
- **.gitignore**: 
  - Git ignore configuration file

- **README.md**  
  This overview, plus usage and license details.

## Usage Example
```python
from datasets import load_dataset

ds = load_dataset("your-username/foodlogathl")
print(ds)
# ➞ DatasetDict({
#       train: Dataset(...)
#   })
# Access:
images      = ds["train"]["images"]
annotations = ds["train"]["annotations"]
metadata    = ds["train"]["metadata"]
```

## User Instructions for FoodLogAthl-218

This dataset is available **exclusively to academic institutions** for **non-commercial research and educational use**.

To request access, please fill out the application form with the required information and click **Apply**.  
Once we review your request (typically within 2–3 business days), you will receive an email notification with access instructions.

---

## After You Apply

- **Review period**: 2–3 business days  
- **Eligibility**: Only applications from academic or research institutions with non-commercial intent will be considered  
- **Contact**: For any questions about access, please email foodlog-hal@hal.t.u-tokyo.ac.jp  
- **Access granted**: If approved, you'll receive read-only access to the `images/` and `annotations/` folders via your Hugging Face account

## Usage Policy

This dataset is strictly limited to academic institutions and may only be used for non-commercial research and educational purposes. 
Commercial use, redistribution, or use by non-academic organizations is **not permitted**. All access requests are manually reviewed to ensure compliance with these conditions.

## References

### ACMMM 2025 Presentation Video

<video controls width="100%">
  <source src="https://huggingface.co/datasets/FoodLog/FoodLogAthl-218/resolve/main/ACMMM_MPV.mp4" type="video/mp4">
  Your browser does not support the video tag.
</video>

**ACMMM 2025 Dataset Track Paper**: This video accompanies our dataset track paper presented at ACMMM 2025 (ACM International Conference on Multimedia), held in Dublin, Ireland, October 28-31, 2025. The paper introduces the FoodLogAthl-218 dataset and provides a comprehensive analysis of real-world meal image recognition challenges.

**DOI**: [10.1145/3746027.3758276](https://doi.org/10.1145/3746027.3758276)