Occlusion Data Augmentation for Object Detectors based on Random Erasing and Bounding Box Tailoring

Recent convolutional object detectors have showed remarkable improvements, but occlusion has been a major obstacle to apply them to the real-world. More specifically, in military domain, cover and concealment of objects are common, and building occlusion-tolerable object detector is important. In this paper, we introduce bounding box tailoring approach based on bivariate standard normal distribution. The main idea of our approach is tailoring bounding boxes to fit tighter to remaining object area after random erasing. Our experiments on military vehicle data set show that our approach outperforms previous random erasing data augmentation techniques.