Scholarlink

<p>This study addresses the challenge of planogram compliance in convenience stores by proposing a scalable, automated shelf monitoring system deployed across over 7,000 7-Eleven stores in Taiwan. Traditional manual audits are labor-intensive, error-prone, and costly, creating a growing need for reliable, automated solutions. To address this challenge, the proposed system integrates computer vision and deep learning techniques into a unified pipeline capable of detecting shelves, recognizing products, and comparing shelf layouts against digital planograms through a customized alignment algorithm. The system further incorporates multi-image stitching to overcome spatial constraints and construct virtual shelves that closely replicate real-world environments, improving adaptability and accuracy. Three large-scale datasets were developed to support model training and validation: 15,232 images for shelf detection, 99,135 images for product detection, and 471 product categories averaging 210 images each for classification. Automated labeling and clustering processes were introduced to substantially reduce manual annotation time. Experimental results demonstrate that the YOLOv8-based detection models achieve exceptional precision and recall across all stages. For shelf detection, the model achieved 99.23% precision, 98.93% recall, and 99.41% mAP@50, while product detection reached 94.61% precision, 93.02% recall, and 95.7% mAP@50—both surpassing transformer-based alternatives such as Deformable DETR. ResNet101 and FAN-based Transformer models achieved 99.86% accuracy on real-world retail datasets, indicating strong model stability. In the few-shot experiments, the FAN-based model showed strong adaptability and generalization, maintaining high accuracy with only five samples per class and achieving 98.39% Top-1 and 99.48% Top-5 accuracy on unseen products, demonstrating excellent transfer learning and real-time recognition capability. The system offers high accuracy, scalability, and real-time efficiency, making it a strong alternative to manual audits and a driver of smart retail innovation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time retail planogram compliance application using computer vision and virtual shelves

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月16日

查看全文
  • Tsung-Yin Ou,
  • Andrés Ponce,
  • Cody Lee,
  • Areoll Wu

摘要

This study addresses the challenge of planogram compliance in convenience stores by proposing a scalable, automated shelf monitoring system deployed across over 7,000 7-Eleven stores in Taiwan. Traditional manual audits are labor-intensive, error-prone, and costly, creating a growing need for reliable, automated solutions. To address this challenge, the proposed system integrates computer vision and deep learning techniques into a unified pipeline capable of detecting shelves, recognizing products, and comparing shelf layouts against digital planograms through a customized alignment algorithm. The system further incorporates multi-image stitching to overcome spatial constraints and construct virtual shelves that closely replicate real-world environments, improving adaptability and accuracy. Three large-scale datasets were developed to support model training and validation: 15,232 images for shelf detection, 99,135 images for product detection, and 471 product categories averaging 210 images each for classification. Automated labeling and clustering processes were introduced to substantially reduce manual annotation time. Experimental results demonstrate that the YOLOv8-based detection models achieve exceptional precision and recall across all stages. For shelf detection, the model achieved 99.23% precision, 98.93% recall, and 99.41% mAP@50, while product detection reached 94.61% precision, 93.02% recall, and 95.7% mAP@50—both surpassing transformer-based alternatives such as Deformable DETR. ResNet101 and FAN-based Transformer models achieved 99.86% accuracy on real-world retail datasets, indicating strong model stability. In the few-shot experiments, the FAN-based model showed strong adaptability and generalization, maintaining high accuracy with only five samples per class and achieving 98.39% Top-1 and 99.48% Top-5 accuracy on unseen products, demonstrating excellent transfer learning and real-time recognition capability. The system offers high accuracy, scalability, and real-time efficiency, making it a strong alternative to manual audits and a driver of smart retail innovation.