Road lane detection using neural networks for the self- driving cars

Semantic segmentation based on convolutional neural networks, used in image regional pixel-wise classification, is a very important tool in computer vision nowadays. It is especially relevant in the development of the software for self-driving cars, who must quickly and accurately identify many objects, such as the road, other cars, pedestrians, etc. However, models performing semantic segmentation require lots of data to achieve desired accuracy, and manually labeling the segmentation data is a very time-consuming job. But now there is a way to efficiently generate decent training data using the computer graphics applications, such as games. This synthetic data can be photo-realistic, and the labels for this data can be efficiently generated automatically. This paper analyzes the application of synthetic data in training of a road detection semantic segmentation model – how this improves the accuracy of the model, when using different amounts of synthetic data and various synthetic/real data ratios. Also, we analyze how to efficiently choose the model parameters, to improve the speed of the model execution, while minimally affecting the accuracy