valiantlynxz/norwegian-grocery-detector

NorgesGruppen Grocery Detection Models

ONNX models for the NorgesGruppen grocery shelf detection competition. Detect and classify 356 product categories on store shelf images.

Quick Start

from huggingface_hub import snapshot_download

# Download the ensemble (best accuracy)
path = snapshot_download(
    "valiantlynxz/norwegian-grocery-detector",
    allow_patterns="submission_ensemble/*",
)

# Download just the lightweight model
path = snapshot_download(
    "valiantlynxz/norwegian-grocery-detector",
    allow_patterns="submission_yolo12n/*",
)

# Download a specific single model
path = snapshot_download(
    "valiantlynxz/norwegian-grocery-detector",
    allow_patterns="submission_yolo12x/*",
)

# Download everything
path = snapshot_download("valiantlynxz/norwegian-grocery-detector")

Variants

Variant	Folder	Architecture	Size	Notes
YOLOv8m	submission/	YOLOv8m	100 MB	General purpose
YOLOv8l	submission_large/	YOLOv8l	168 MB	Higher accuracy
YOLO12n	submission_yolo12n/	YOLO12n	11 MB	Lightweight/fast
YOLO12x	submission_yolo12x/	YOLO12x	228 MB	Best single model
Ensemble	submission_ensemble/	WBF (3 models)	407 MB	Best overall accuracy

Ensemble Details

submission_ensemble/ fuses predictions from 3 models using Weighted Boxes Fusion:

Model file	Architecture	Weight	Size
model_yolo12x.onnx	YOLO12x	3	228 MB
model_yolov8l.onnx	YOLOv8l	2	168 MB
model_yolo12n.onnx	YOLO12n	1	11 MB

Inference

Each submission folder contains a self-contained run.py:

python submission_ensemble/run.py --input /path/to/images --output predictions.json

Output format (COCO):

[{"image_id": 1, "category_id": 42, "bbox": [141, 49, 169, 152], "score": 0.95}]

Programmatic Variant Selection

import json
from pathlib import Path
from huggingface_hub import hf_hub_download, snapshot_download

# Load config to discover variants
config_path = hf_hub_download("valiantlynxz/norwegian-grocery-detector", "config.json")
config = json.loads(Path(config_path).read_text())

# List available variants
for name, info in config["variants"].items():
    print(f"{name}: {info['architecture']} ({info['size_mb']} MB) -> {info['folder']}/")

# Download a specific variant by name
variant = config["variants"]["yolo12x"]
path = snapshot_download(
    "valiantlynxz/norwegian-grocery-detector",
    allow_patterns=f"{variant['folder']}/*",
)

Requirements

• Python 3.11
• onnxruntime-gpu (CUDA 12.4) or onnxruntime (CPU)
• opencv-python, numpy
• ensemble_boxes (ensemble variant only)

Training

Trained on valiantlynxz/norwegian-grocery:

• 248 shelf images, 22,731 annotations, 356 categories
• Input size: 1280x1280
• Augmentations via ultralytics + albumentations

Usage with the Detection Project in https://github.com/kuben-labs/nmai

cd detection/
make models    # downloads all variants
make setup     # downloads data + models + generates YOLO format