openfoodfacts/price-tag-detection

Open Prices - Price tag detection model

Wandb tracking run

This object detection model was trained on images from the Open Prices database to detect price tags of products in shops.

It was trained on 200 epochs using Ultralytics library, using yolov11x version, with images resized to 960x960. This model is licensed under the AGPLv3 license.

Model Details

Model Description

• Developed by: Open Food Facts
• Model type: object detection
• License: agpl-3.0
• Finetuned from model: yolo11x.pt

Training Details

Training Data

The model was fine-tuned using the following dataset: openfoodfacts/price-tag-detection (revision: v1.2).

Training Procedure

Dependency versions:

• ultralytics: 8.4.14
• pytorch: 2.9.0+cu128

Training Hyperparameters

• Epochs: 200
• Batch size: 16
• Image size: 960

Evaluation

The following evaluation metrics were obtained after training the model:

• metrics/precision(B): 0.9405504116589396

• metrics/recall(B): 0.8999179655455292

• metrics/mAP50(B): 0.9490994303327694

• metrics/mAP50-95(B): 0.8022528188009798

• fitness: 0.8022528188009798

Evaluation on exported models

The model was also evaluated after exporting to ONNX and TensorRT formats. The following metrics were obtained:

ONNX export

• metrics/precision(B): 0.9393607953718663

• metrics/recall(B): 0.9048400328137818

• metrics/mAP50(B): 0.9485645846101123

• metrics/mAP50-95(B): 0.8000803959473284

• fitness: 0.8000803959473284

Files

Most files stored on the repo are standard files created during training with the Ultralytics YOLO library.

What was added:

• an ONNX export of the trained model (best model), stored in weights/model.onnx.
• a Parquet file containing predictions on the full dataset, stored in predictions.parquet.
• metrics JSON files for each exported model format, stored in metrics_*.json:
  • metrics.json: metrics for the original PyTorch model
  • metrics_onnx.json: metrics for the ONNX exported model

History

• v2.0: improved version of the model, see the report.
• v1.0: first release of the model.