SohlHealth/sohl-multidish-yolo-dataset

🍽️ SOHL Multi-Dish Indian Food Detection Dataset

Overview

This dataset contains 377 annotated images of Indian food plates with multiple dishes per image. Designed for training YOLO models to detect and classify multiple food items on a single plate.

Dataset Statistics

• Images: 377
• Annotations: 377
• Classes: 16
• Format: YOLOv8 (images + txt annotations)
• Created: 2025-08-16

Classes

0. bread_or_Roti_naan - Chapati, naan, roti, paratha, and other Indian breads
1. curry_dish - General curry preparations, gravies, and liquid dishes
2. rice_dish - Plain rice, biryani, pulao, and rice preparations
3. dry_vegetable - Bhindi, aloo, cauliflower, and dry sabzi preparations
4. snack_item - Samosa, pakora, vada, dhokla, and fried snacks
5. sweet_item - Traditional sweets, desserts, and mithai
6. accompaniment - Pickle, raita, papad, chutney, and side dishes
7. Dal_or_sambar - Dal preparations, sambar, and lentil-based dishes
8. drink - Beverages, juices, lassi, and liquid refreshments
9. eggs - Egg preparations, omelettes, and egg-based dishes
10. fish_dish - Fish curry, fried fish, and seafood preparations
11. fruits - Fresh fruits, fruit salads, and fruit-based items
12. pasta - Pasta dishes and Italian preparations
13. salad - Vegetable salads, mixed salads, and fresh preparations
14. soup - Soups, broths, and liquid appetizers
15. south_indian_breakfast - Dosa, idli, upma, and South Indian breakfast items

Dataset Structure

sohl-multidish-yolo-dataset/
├── images/           # 377 image files
├── labels/           # 377 YOLO format annotations
├── dataset.yaml      # YOLOv8 configuration
└── README.md         # This file

Usage

Download Dataset

from huggingface_hub import snapshot_download

# Download entire dataset
dataset_path = snapshot_download(
    repo_id="SohlHealth/sohl-multidish-yolo-dataset",
    repo_type="dataset"
)

Train YOLOv8

from ultralytics import YOLO

# Load model and train
model = YOLO('yolov8s.pt')
results = model.train(
    data='dataset.yaml',
    epochs=100,
    batch=8,
    imgsz=640
)

Key Features

• ✅ Multi-dish detection: 2-6 items per plate
• ✅ Indian cuisine focus: Traditional dishes and combinations
• ✅ Real-world scenarios: Restaurant and home environments
• ✅ Complex layouts: Overlapping items, various plate styles
• ✅ High-quality annotations: Precise bounding boxes
• ✅ Comprehensive classes: 16 food categories including regional specialties

Performance Expectations

Based on similar datasets and architectures:

• Expected mAP@0.5: 15-25% (multi-dish detection is challenging)
• Training time: 3-6 hours on modern GPU
• Recommended epochs: 100-150
• Best practices: Transfer learning from food detection models

Citation

@dataset{sohl_multidish_dataset_20250816_161951,
  title={SOHL Multi-Dish Indian Food Detection Dataset},
  author={SOHL AI Team},
  year={2025},
  url={https://huggingface.co/datasets/SohlHealth/sohl-multidish-yolo-dataset}
}

License

MIT License - See LICENSE file for details.

Contact

For questions about this dataset, please contact the SOHL AI team.