Product object detection using pseudo-generated product shelf images

Abstract

In the retail industry, there is a growing need for AI to identify products on product shelves in order to improve the efficiency of product management and planogram. One of the problems in training a product object detection model using product shelf images of actual stores is the labor and cost required to collect and annotate a large number of product shelf images. In this research, we built a system that automatically generates pseudo product shelf images that are close to the real thing by synthesizing images of the parts that make up the product shelf, using prior knowledge about the display characteristics of each product type, obstacles such as promotional items, and variations in shooting quality due to lighting, angle of view, and camera shake. As a result of training the state-of-the-art object detection model on pseudo product shelf images, the detection performance in product shelf images of actual stores achieved AP=96%, realizing practical level performance.