Training Image Synthesis for Shelf Item Detection reflecting Alignments of Items in Real Image Dataset

We propose a novel cut-and-paste approach to synthesize a training dataset for shelf item detection, reflecting the alignments of items in the real image dataset. The conventional cut-and-paste approach synthesizes large numbers of training images by pasting foregrounds on background images and is effective for training object detection. However, the previous method pastes foregrounds on random positions of the background, so the alignment of items on shelves is not reflected, and unrealistic images are generated. Generating realistic images that reflect actual positional relationships between items is necessary for efficient learning of item detection. The proposed method determines the pasting positions for the foreground images by referring to the alignment of the items in the real image dataset, so it can generate more realistic images that reflect the alignment of the real-world items. Since our method can synthesize more realistic images, the trained models can perform better.