Synthetic Data Generation in Virtual Environments for Automated Shelf Auditing in Supermarkets

The automation of shelf auditing in retail environments depends on robust computer vision models for product detection and classification, yet the limited availability of labeled datasets poses a major challenge for training accurate and adaptable models. This work addresses this issue by proposing a synthetic data generation framework based on virtual environments, leveraging 3D product models and realistic supermarket shelf simulations to create diverse and scalable datasets. Two datasets are introduced: RealSynthDet-15K, designed for product detection by combining 11,739 real images from SKU-110K with 3,500 synthetic images, and ProdSynth-20, a classification dataset containing 50,000 images across 20 product categories. The methodology integrates automated product placement, lighting variability, and viewpoint adjustments to generate high-fidelity training data, which is then used to train a detection and classification pipeline with YOLOv8. Experimental results demonstrate a well-balanced trade-off between precision and recall, confirming the effectiveness of synthetic datasets in real-world applications while reducing reliance on manual data collection. This study highlights the potential of virtual environments as a scalable and cost-effective alternative for dataset creation in retail automation, paving the way for future improvements in 3D modeling fidelity and the exploration of advanced deep learning architectures to further optimize performance in commercial settings.