Аннотация:Modern computer vision methods usually require lots of labelled data for training. Besides price of labelling, problems with rare object classes and adaptation to new domain or task arise. One of the promising methods to solve these problems is to generate synthetic training data. In this work we focus on task of traffic sign detection. We consider several methods for generating synthetic data for training traffic sign detectors: random placement of signs of different quality (simple synthetic, CGI based and CGI improved using generative adversarial network). We also propose a method to replace real signs with synthetic signs. Experimental evaluation shows that proposed method improves quality of detection of rare traffic signs and that usage of synthetic data is very helpful for improving training of traffic sign classifier.