Improving Automatic Check-Out Accuracy with Fine-tuned YOLOv10 and YOLOv12 Models

Improving the retail customer experience includes solving the Automatic Check-Out (ACO) problem, which involves predicting a receipt based on a photograph of products in a checkout area. Recent work demonstrated that the ACO problem can be addressed by fine-tuning a strong pre-trained object detection model from the You Only Look Once (YOLO) series on a limited set of checkout photographs. Our work extends this research by introducing an additional filtering step for the results from YOLOv10 and YOLOv12 models. Furthermore, we investigate improving checkout accuracy by merging detections from an ensemble (pair) of different YOLO variants. The proposed approach can achieve a competitive checkout accuracy of up to 96.89% on the RPC dataset.