README.md

metadata

dataset_info:
  features:
    - name: image_id
      dtype: string
    - name: image
      dtype: image
    - name: width
      dtype: int64
    - name: height
      dtype: int64
    - name: meta
      struct:
        - name: barcode
          dtype: string
        - name: off_image_id
          dtype: string
        - name: image_url
          dtype: string
    - name: objects
      struct:
        - name: bbox
          sequence:
            sequence: float32
        - name: category_id
          sequence: int64
        - name: category_name
          sequence: string
  splits:
    - name: train
      num_bytes: 576037866.625
      num_examples: 1083
    - name: val
      num_bytes: 64207578
      num_examples: 123
  download_size: 1304859691
  dataset_size: 640245444.625
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*

Open Food Facts Nutrition table detection dataset

This dataset was used to train the nutrition table object detection model running in production at Open Food Facts.

Images were collected from the Open Food Facts database and labeled manually. Just like the original images, the images in this dataset are licensed under the Creative Commons Attribution Share Alike license (CC-BY-SA 3.0).

Fields

• image_id: Unique identifier for the image, generated from the barcode and the image number.
• image: Image data.
• width: Image original width in pixels.
• height: Image original height in pixels.
• meta: Additional metadata.
  • barcode: Product barcode.
  • off_image_id: Open Food Facts image number.
  • image_url: URL to the image on the Open Food Facts website.
• objects: Object detection annotations.
  • bbox: List of bounding boxes in the format (y_min, x_min, y_max, x_max). Coordinates are normalized between 0 and 1, using the top-left corner as the origin.
  • category_id: List of category IDs.
  • category_name: List of category names.

Versions

• 1.0: Original data used to train the tf-nutrition-table-1.0 model.
• 1.1: Fixes erroneous bounding boxes due to rotated images. About 10% of the bounding boxes were completely wrong due to the annotation software we were using. All samples were manually reviewed again, and the erroneous bounding boxes were corrected. Also, ~35 duplicated images were removed from the dataset (images that belong to the same product).