Synthetic dataset generation system for vehicle detection

The success of machine learning models for object detection highly depends on the training data size and quality. Generating synthetic data speeds up the data acquisition process by removing the need for human annotation. Moreover, since annotation is done automatically, there is no room for human error. We present a pipeline that automatically generates and annotates aerial images of vehicles on roads. The pipeline is structured to allow easy adding of various new vehicles and is not limited to cars only. The resolution of the generated images and the level of detail can be modified by changing the output settings. • The pipeline can be used for fast creation of large scale annotated data that is needed for training machine learning models. • Aerial images of cars created with this pipeline are realistic which is important to achieve when generating synthetic data. • The structure of the pipeline allows easy adding of new vehicle types and creation of custom datasets. • This is a baseline for automated annotated data creation using 3D modeling software, customizable for various use cases.