NorgesGruppen Grocery Shelf Detection

Object detection model for identifying grocery products on store shelves. Fine-tunes YOLOv8x on COCO-format shelf annotations across 356 product categories.

Setup

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies (pinned to match sandbox versions)
pip install -r requirements.txt

For GPU training with CUDA:

pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
pip install ultralytics==8.1.0

Data Setup

Place the downloaded datasets in data/:

data/
├── coco/
│   ├── images/           ← unzip NM_NGD_coco_dataset.zip here
│   │   ├── img_00001.jpg
│   │   └── ...
│   └── annotations.json
└── product_images/       ← unzip NM_NGD_product_images.zip here
    ├── 7040913336691/
    │   ├── main.jpg
    │   ├── front.jpg
    │   └── ...
    └── metadata.json

Workflow

1. Explore the data

python src/explore_data.py --coco_dir data/coco --product_dir data/product_images --top_n 20

Shows the rarest product categories with shelf crops and reference photos.

2. Convert COCO to YOLO format

python src/convert_coco_to_yolo.py --coco_dir data/coco --output_dir data/yolo

Creates data/yolo/ with train/val split and generates config.yaml.

3. Train

python src/train.py --epochs 100 --batch 8 --imgsz 640

Trains YOLOv8x with COCO-pretrained weights. Best model saved to runs/detect/train/weights/best.pt.

4. Evaluate locally

python src/evaluate.py --model runs/detect/train/weights/best.pt --save_images

Reports the competition-style combined score: 0.7 * detection_mAP + 0.3 * classification_mAP.

5. Package submission

python src/package_submission.py --weights runs/detect/train/weights/best.pt

Creates submission.zip with run.py and best.pt at the root. Validates size and file constraints.

6. Submit

Upload submission.zip on the competition submit page. The sandbox runs:

python run.py --input /data/images --output /output/predictions.json

Sandbox Environment

Resource	Value
GPU	NVIDIA L4 (24 GB VRAM)
Python	3.11
ultralytics	8.1.0
torch	2.6.0+cu124
Timeout	300 seconds
Network	None (fully offline)
Max zip	420 MB uncompressed

Scoring

Score = 0.7 × detection_mAP@0.5 + 0.3 × classification_mAP@0.5

• Detection (70%): did you find the products? (IoU >= 0.5, category ignored)
• Classification (30%): did you identify the right product? (IoU >= 0.5 AND correct category_id)