Composite Image Generation Using Labeled Segments for Pattern-Rich Dataset without Unannotated Target

Although object detection technology using cameras offers potential for various applications, it incurs dataset creation costs to train new models where general-purpose models are ineffective, such as in industrial settings. We have previously developed a semi-automated annotation framework that employs optical flow and representation learning techniques to reduce human effort significantly. However, it was likely to cause unintended annotation omissions and mistakes compared to manual annotation. In this study, we propose a composite image generation approach to create omission-free and pattern-rich datasets. The proposed method synthesizes natural-looking images without unannotated targets by placing labeled foreground segments at their original positions on targetless background frames collected with the same fixed-point cameras. Evaluation with video footage in a logistics warehouse confirmed that improved dataset reliability led to higher model performance.