Enhancing Object Detection: A Data Augmentation Approach Using Rendered Foreground Data and Image Postprocessing
Jie Qing Fan, Yongqiang Xie, Zhongbo Li · 2025
Inserting rendered foreground into real images is a common method for generating synthesized data. The same way, inserting rendered foreground into datasets that captured in real world could be an effect data augmentation method. In this paper, we propose mosaic instance insertion for data augmentation (Mii4da) to expand dataset by inserting rendered foregrounds into backgrounds from the target dataset. We also perform some additional postprocessing by mimicking the feature of images in target dataset to bridge the domain gap between foreground and the background. We demonstrate the effectiveness of Mii4da in 2d vehicle detection tasks, surpasses other synthetic data generation approaches, such as VKITTI2, Sim10k, as well as real-world datasets, like BDD100K and Cityscapes. Furthermore, combining synthetic Mii4da with real KITTI data yields superior results compared to using real KITTI data only.