The Influence of Transfer Learning on the Accuracy of Object Localization and Classification in the YOLO Detector

In this paper the influence of transfer learning on the accuracy of classification and localization of the YOLO detector in the problem of transport traffic analysis is studied. The authors research the possibility of adaptation of neural network model to specific scenarios. Adaptation is supposed to be carried out by transfer learning with freezing of some network layers. In the paper the dependence of the model accuracy in the process of transfer learning by specifying the number of frozen layers is shown. Several YOLOv5 models have been studied: «nano», «small», and «medium». Experimental research has been carried out on new custom dataset, and conclusion has been made about the effectiveness of transfer learning for vehicle classification and localization based on performance metrics.