Experimental Results on Synthetic Data Generation in Unreal Engine 5 for Real-World Object Detection

This paper investigates a new method for generating high-quality synthetic training data for object detection (OD) algorithms. Utilizing photogrammetry, we transform real-world objects into 3D digital replicas in Unreal Engine 5 (UE5). We then simulate diverse scenarios, capturing images of these objects at varying angles, backgrounds, and textures. To evaluate the data quality, we experiment with multiple configurations and train OD models using YOLOv8. Our findings highlight the effectiveness of the proposed image generation approach in UE5 and establish its potential as a cost-efficient solution for generating extensive volumes of training data for OD algorithms.