Solving automatic check-out with fine-tuned YOLO models
Rafał Skinderowicz · Procedia Computer Science · 2024
The Automatic Check-Out (ACO) problem entails the automatic detection and classification of products within a checkout image. Contrary to conventional methods that depend on synthetic images for training, this paper introduces a methodology that involves fine-tuning pretrained YOLOv8 and YOLOv9 models using a selected subset of validation images from the Large-Scale Retail Product Checkout (RPC) dataset. We demonstrate that the fine-tuning outcomes can be enhanced through the generation of additional examples by merging pairs of existing checkout images. Further enhancement of the models’ accuracy can be achieved by implementing a modified non-maximum suppression algorithm to sift through the detection results. The employed approach yields an average checkout accuracy of 95.16% on the test segment of the RPC dataset, showcasing the efficacy of leveraging pretrained models to address the complexities of real-world image recognition challenges in retail settings.