Object Detection for Autonomous Driving in Chinese Rural Scenes: An Experimental Study on Real-Synthetic Data Mixing and Model Evaluation

Autonomous driving perception systems are primarily trained on urban-centric datasets, leaving them ill-equipped for the unique traffic participants and infrastructure of rural China. This study investigates how to bridge this data gap by constructing a specialized real-synthetic mixed dataset and evaluating how different ratios of synthetic data impact the performance of 13 state-of-the-art object detectors.

The researchers captured 4,720 real-world images from Weishi County, Henan Province, focusing on rural-specific elements like electric tricycles, low-speed vehicles (LSVs), and roadside stalls. They complemented this with a synthetic dataset generated via Unreal Engine, using high-fidelity assets to replicate rural road conditions. They then systematically trained 13 models—including the YOLOv5, v8, 11, 26 series, and RT-DETR-L—across three configurations: an all-real baseline, a 1:0.5 real-to-synthetic mix, and a 1:1 mix. All models were evaluated under a unified training protocol to ensure comparability.

The study reveals that a 1:0.5 real-to-synthetic data ratio acts as an optimal "sweet spot," effectively boosting detection performance with YOLO11m achieving the highest mAP@0.5 of 0.758. Increasing the synthetic data proportion to 1:1, however, triggers domain shifts—where the synthetic data's visual characteristics diverge enough from reality to offset the benefits of increased training volume. While most models successfully identify standard vehicles, the researchers identified a persistent "perceptual bottleneck" for long-tail, non-standard objects such as roadside stalls and railings, suggesting that simply adding more synthetic data is insufficient to solve these specific detection challenges.

This research provides a practical roadmap for developers deploying autonomous systems in rural regions. By demonstrating that synthetic data is a powerful but sensitive tool, the study offers empirical evidence that model performance is not strictly correlated with data quantity. These insights help reduce the high costs of manual annotation while highlighting the specific architectural and data-mixing strategies required to handle the complex, unstructured nature of rural traffic.