Improving Detection of Similarly Designed Retail Products With Size Variants

Abstract

Retail environments require accurate and consistent shelf monitoring to support inventory management and planogram compliance. However, manual inspection remains inefficient and error-prone, particularly in modern grocery stores that contain thousands of Stock Keeping Units (SKUs). As a result, automated vision-based monitoring systems have been increasingly adopted to replace manual shelf inspection and improve scalability. Moreover, a persistent challenge in automated retail monitoring systems arises when product variants from the same brand share nearly identical packaging designs and differ only in weight or grammage. In such fine-grained scenarios, object detection models frequently confuse product variants, leading to

unreliable inventory data and missed detections. This research addresses the problem of fine- grained retail product detection under real-world constraints, where data scarcity and class

imbalance limit the effectiveness of deep learning models. To mitigate these challenges, an optimization-driven detection pipeline is proposed, integrating systematic hyperparameter tuning, dataset-level context-aware augmentation, and ensemble model post-processing. Two complementary object detection architectures, YOLOv12 and Faster R-CNN are employed to capture diverse feature representations. Model training is optimized using hierarchical hyperparameter tuning with the Optuna Tree-Structured Parzen Estimator (TPE), while data limitations are addressed through Context-Aware Copy-Paste augmentation. Furthermore, prediction outputs are fused using Weighted Boxes Fusion (WBF) to improve detection consistency. Experimental evaluation on real-world retail shelf images demonstrates that the proposed pipeline reduces error rates compared to baseline detection configurations. Analysis using confusion matrices shows a reduction in misclassification among product variants. These findings indicate that combining architecture diversity, systematic optimization, data-centric augmentation, and ensemble model provides a solution for fine-grained retail product detection in data-limited retail environments.