NorgesGruppen Data: Object Detection

Detect grocery products on store shelves. Upload your model code as a .zip file — it runs in a sandboxed Docker container on our servers.

How It Works

Download the training data from the competition website (requires login)
Train your object detection model locally
Write a run.py that takes shelf images as input and outputs predictions
Zip your code + model weights
Upload at the submit page
Our server runs your code in a sandbox with GPU (NVIDIA L4, 24 GB VRAM) — no network access
Your predictions are scored: 70% detection (did you find products?) + 30% classification (did you identify the right product?)
Score appears on the leaderboard

Downloads

Download training data and product reference images from the Submit page on the competition website (login required).

Training Data

Two files are available for download:

COCO Dataset (NM_NGD_coco_dataset.zip, ~864 MB)

248 shelf images from Norwegian grocery stores
~22,700 COCO-format bounding box annotations
356 product categories (category_id 0-355) — detect and identify grocery products
Images from 4 store sections: Egg, Frokost, Knekkebrod, Varmedrikker

Product Reference Images (NM_NGD_product_images.zip, ~60 MB)

327 individual products with multi-angle photos (main, front, back, left, right, top, bottom)
Organized by barcode: {product_code}/main.jpg, {product_code}/front.jpg, etc.
Includes metadata.json with product names and annotation counts

Annotation Format

The COCO annotations file (annotations.json) contains:

{
  "images": [
    {"id": 1, "file_name": "img_00001.jpg", "width": 2000, "height": 1500}
  ],
  "categories": [
    {"id": 0, "name": "VESTLANDSLEFSA TØRRE 10STK 360G", "supercategory": "product"},
    {"id": 1, "name": "COFFEE MATE 180G NESTLE", "supercategory": "product"},
    ...
    {"id": 356, "name": "unknown_product", "supercategory": "product"}
  ],
  "annotations": [
    {
      "id": 1,
      "image_id": 1,
      "category_id": 42,
      "bbox": [141, 49, 169, 152],
      "area": 25688,
      "iscrowd": 0,
      "product_code": "8445291513365",
      "product_name": "NESCAFE VANILLA LATTE 136G NESTLE",
      "corrected": true
    }
  ]
}

Key fields: bbox is [x, y, width, height] in pixels (COCO format). product_code is the barcode. corrected indicates manually verified annotations.

What Annotations Look Like

A training image with all ground truth boxes (green = correctly detected product):

Shelf image with all 76 products annotated in green bounding boxes — this represents a perfect mAP of 1.0

Compare with a ~50% mAP result — half the products are missed entirely, and some detected boxes (red) are imprecise:

Same shelf image with only half the products detected, some with imprecise boxes shown in red — approximately 50% mAP

Submit

Upload your .zip at the submission page on the competition website.

MCP Setup

Connect this docs server to your AI coding tool:

claude mcp add --transport http nmiai https://mcp-docs.ainm.no/mcp