Improving the Efficiency of Classifier of Hand Disinfection Stages Using Synthesised Video Sequences

Thanks to the progress in 3D graphics, it has become possible to train models on synthetic images, which can help to get rid of the need to collect and label data. However, training on generated images can lead to lower model quality due to the difference between synthetic and real image distributions. To solve this problem, a method is proposed that improves the realism of the output of the graphical simulator with little real data. It uses a conditional generative adversarial network (cGAN) with synthetic images as input instead of random vectors, with a few changes to preserve image content, avoid artifacts and stabilize learning. The work shows that this allows you to create realistic images, which are later used to train the classifier of the stages of hand disinfection, and to improve the classification results compared to the use of synthetic images.