A Real-Synthetic Mixed Object Detection Dataset for Autonomous Driving in Chinese Rural Scenes

Engineering notes

Intended Use Training and benchmarking object detectors (e.g., YOLO series, RT‑DETR) for autonomous driving perception.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

1. Overview This dataset is constructed for object detection in autonomous driving scenarios, specifically targeting the complex and under‑represented rural traffic environments in China. It comprises real rural road images collected in Henan Province and high‑fidelity synthetic images generated using Unreal Engine 5.7. The dataset is designed to address the data scarcity and domain generalization challenges faced by perception models when deployed in rural areas. 2. Data Composition Real images: 4,720 frames extracted from on‑road video recordings (February 2026) covering three typical rural scene types: county roads, market town centers, and village streets. Synthetic images: 4,200 photorealistic virtual images rendered with Unreal Engine, featuring parametrically controlled road layouts, object placements, lighting conditions, and weather variations. Total samples: Up to 8,400 images across different configurations (the full set includes both real and synthetic subsets). All images are provided in JPG/PNG format with corresponding YOLO‑format annotation files (.txt). 3. Object Categories A fine‑grained 14‑category taxonomy is defined to cover the unique traffic elements in Chinese rural scenes: ID Category Remarks 0 scooter Primary individual transport tool 1 tricycle Core short‑distance freight/passenger vehicle 2 LSV Low‑speed electric vehicle (elderly mobility) 3 stall Roadside temporary vendor, key obstacle 4 bin Common roadside trash container 5 person Pedestrian (often not on sidewalks) 6 car Sedan, SUV, MPV 7 truck Freight trucks and semi‑trailers 8 billboard Visual distractor 9 railing Road boundary indicator 10 pole Utility poles, sign poles 11 street lamp Roadside lighting 12 sign Traffic/road signs (diverse styles) 13 tree Main roadside vegetation 4. Dataset Splits The dataset is organized into three configurations for controlled experiments: Config A: Real‑only (4,200 images) Config B: Real : Synthetic = 1 : 0.5 (4,200 real + 2,100 synthetic) Config C: Real : Synthetic = 1 : 1 (4,200 real + 4,200 synthetic) A fixed test set of 517 real images (covering all categories and diverse environments) is provided separately for unbiased evaluation. 5. Annotation Format All annotations are in YOLO Darknet format (class_id x_center y_center width height), normalized to image dimensions. Bounding boxes were manually annotated for real images using a two‑round cross‑validation process, and automatically generated for synthetic images via semantic segmentation to YOLO conversion scripts. 6. Intended Use Training and benchmarking object detectors (e.g., YOLO series, RT‑DETR) for autonomous driving perception. Studying the effects of real‑to‑synthetic data mixing ratios on model performance. Evaluating domain adaptation and generalization in unstructured rural traffic scenes.

Links and sources

Comments