MIST: Multi-Domain Synthetic Dataset for Rural Driving

Abstract

Existing driving datasets are largely urban-centric, predominantly collected under daytime and clear weather conditions. As a result, rural driving scenarios with sparse traffic, long-range road structures, and pronounced environmental variations remain underrepresented, despite their importance in real-world driving. This bias limits systematic analysis of domain shifts and hinders the development of models robust to environmental changes beyond urban settings. In this work, we introduce MIST, a Multi-Domain Synthetic Dataset for Rural Driving, constructed to address the lack of rural driving data. The dataset is built from a simulation environment and designed with explicitly structured domain factors such as season, time of day, and weather. By combining these factors, MIST provides 32 domain configurations with an equal number of samples for each configuration, totaling approximately 200K images. Each image is accompanied by instance segmentation annotations and paired image–text descriptions. This design enables controlled investigation of domain variations. To demonstrate the utility of the proposed dataset, we conduct a series of experiments including zero-shot domain classification and image-to-image translation. The results show that MIST supports diverse downstream analyses and can serve as a practical benchmark for studying domain generalization and data augmentation in rural driving scenarios. The dataset is publicly available via https://huggingface.co/datasets/jongwonryu/MIST_dataset