Abstract The identification and management of grocery items in retail environments have traditionally relied on barcode-based systems, which require significant human intervention and underutilize existing surveillance infrastructure. Computer vision–based approaches offer a promising alternative for automated product recognition. However, many existing grocery datasets remain relatively homogeneous or limited in scale, geographic diversity, or real-world variability. To support more realistic evaluation settings, we present a large-scale grocery dataset collected from eight stores across multiple states in India. The dataset comprises over 13,000 images spanning 349 product categories and captures practical retail challenges such as dense shelf arrangements, occlusions, viewpoint variations, and visual ambiguity. Rather than claiming novelty in addressing these challenges individually, our contribution lies in systematically integrating them within a unified and diverse dataset framework. We also introduce a lightweight product identification pipeline based on omni-scale feature learning, designed to balance representational capacity and computational efficiency. The proposed model achieves a mAP@0.50 of 58.3, a precision of 72.9%, and a recall of 77.9% on the proposed dataset, demonstrating competitive performance while maintaining a compact architecture. Comprehensive comparisons with established benchmark models further contextualize our contributions within the broader literature. Overall, this work provides a diverse evaluation benchmark and an efficient detection framework for practical retail deployment. Similar content being viewed by others Introduction Object detection has remained a core research area within computer vision, with its progress significantly accelerated by advances in deep learning. Over the past decade, the field has transitioned from traditional handcrafted feature-based techniques, such as SIFT1 and HOG2-to highly optimized deep learning architectures. This shift began with the introduction of two-stage detectors, most notably R-CNN3, and later evolved toward faster and more efficient single-stage detection frameworks including YOLO4, YOLOv35, YOLOv56, YOLOv87, YOLOv98, YOLOv109, YOLOv1110, YOLO2611 and RetinaNet12. More recent developments have introduced transformer-based architectures such as