Main Approaches to the Preparation of Visual Data for Training Neural Network Algorithms

— An analysis of the main approaches used by the developers of neural network algorithms for the preparation of training data and the formation of training samples has been carried out. Possible ways of obtaining labeled images have been considered. Examples of open labeled image libraries include ImageNet or Coco offering labeled and annotated photo images, as well as 3D data set libraries. Specialized image marking editors for working with non-standard objects and objects incompletely represented in public libraries, which allow marking up data both manually and in semi-automatic mode, have been studied. Synthetic data generators and simulators have been considered that allow simulating hard-to-reproduce events, as well as a combined approach using GAN-type networks. An analysis has been made of the main difficulties that developers face when preparing training data, including the analysis of the shortcomings of ready data sets, synthetic generators and an approach using GAN-type networks .