Detection of Tea Box Orientations in Retail Shelves Images

This paper presents a new deep learning-based method for the detection of tea box orientations in shelf images. Due to the lack of data, we had to generate synthetic datasets using the Blender rendering framework. This is done by first hand-crafting a 3D model of a retail shelf which is then stocked with a selection of various tea boxes. These 3D models of the tea boxes are obtained by the process of
$\mathbf{3D}$
scanning and uploading the scans to the Blender rendering software. Next, the behavior of customers is simulated using Markov chains, and therefore some tea boxes are removed from the shelf. While stocking the tea boxes they are randomly oriented for the sake of dataset variance maximization. Finally, 10000 virtual retail scenes are rendered and annotated in three different annotation sets. We train the YOLOv5 object detection model on the synthetic datasets and achieve an impressive result of up to 84.2% mean average precision. This experimental result shows high promise for future research in this field of object pose estimation in dense scene images.