Exploring the Potential of Synthetic Data for Bike Path Surface Classification using Diffusion Models

Abstract

Road surface classification of bike paths enables image recognition applications for bike route planning, navigation optimization or path maintenance. However, acquiring and annotating real data can be costly and time-consuming. Synthetic data can overcome data scarcity and annotation costs. We use synthetic data, generated by Stable Diffusion to improve neural network performance on new or unseen surfaces. We compare model performance for different real-synthetic data ratios. Our results show that synthetic data decreases the amount real data needed and improves neural network performance in road surface classification on new surfaces.