Towards Using Synthetic Data in Aerial Image Segmentation

This paper explores the use of synthetic datasets to improve aerial image segmentation, addressing the need for large and diverse data for model training. Current benchmarks often lack real-world conditions, such as high-altitude and nadir perspectives. To overcome this, we propose a controlled data generation approach using the CARLA simulator to generate aerial images of different towns under different weather and time of day conditions, with dynamic traffic elements. We compare our dataset with existing real and synthetic datasets, and evaluate model performance by training the DeepLabV3+ neural network on our dataset and testing on real data. The results show that incorporating synthetic data yields performance comparable to training on real data alone, highlighting its complementary value.