Abstract
Synthetic data has become an increasingly important tool for semantic segmentation, where collecting large-scale annotated datasets is often costly and impractical. Prior work has leveraged computer graphics and game engines to generate training data, but many pipelines remain limited to single modalities and constrained environments or require substantial manual setup. To address these limitations, we present a fully automated pipeline built within Unreal Engine 5 (UE5) that procedurally generates diverse, labeled environments and collects multimodal visual data for semantic segmentation tasks. Our system integrates UE5's biome-based procedural generation framework with a spline-following drone actor capable of capturing both RGB and depth imagery, alongside pixel-perfect semantic segmentation labels. As a proof of concept, we generated a dataset consisting of 1169 samples across two visual modalities and seven semantic classes. The pipeline supports scalable expansion and rapid environment variation, enabling high-throughput synthetic data generation with minimal human intervention. To validate our approach, we trained benchmark computer vision segmentation models on the synthetic dataset and demonstrated their ability to learn meaningful semantic representations. This work highlights the potential of game-engine-based data generation to accelerate research in multimodal perception and provide reproducible, scalable benchmarks for future segmentation models.