DETECTION OF VISUALLY SIMILAR RETAIL PRODUCTS WITH SIZE VARIANTS: A COMPARISON OF YOLOv12 AND FASTER R-CNN

Abstract

Fine-grained retail product detection presents unique challenges due to the high visual similarity among packaging variants and dense shelf arrangements, which often limit the effectiveness of generic detection strategies. This study investigates the baseline performance characteristics of two distinct object detection paradigms—the modern single-stage YOLOv12 and the established two-stage Faster R-CNN—to understand their inherent behaviors in a retail environment. Utilizing a custom dataset of product size variants, both architectures were evaluated under default inference protocols to establish raw performance benchmarks without extensive optimization. The experimental results reveal distinct performance profiles: the best-performing single-stage variant, YOLOv12x, achieved a Mean Average Precision (mAP50:95) of 0.762 with high inference efficiency, while the top-performing two-stage configuration, Faster R-CNN (ResNet50v2), provided a solid structural baseline of 0.709. These findings provide empirical insights into the trade-offs of modern detectors, serving as a foundational reference for developing robust, potentially hybrid, smart retail systems.