SnowRef-Drive: A Global Instruction-Driven Traffic Video Segmentation Dataset for Adverse Winter Driving Scenarios

Adverse weather, particularly snowfall and low visibility, severely degrades vision-based traffic perception, yet existing language-driven video segmentation datasets rarely provide pixel-level annotations for real-world driving under such conditions. We present SnowRef-Drive, a large-scale global dataset for instruction-driven traffic video segmentation in adverse winter environments. The dataset contains 19,700 short clips (118,200 annotated frames), each sampled at 2 FPS over a 3-second duration (6 frames), paired with a natural-language segmentation instruction and frame-wise instance masks. SnowRef-Drive spans 20 winter driving regions across North America, Europe, and East Asia, covering diverse road topologies, including urban roads, highways, mountain/alpine routes, and mixed-transition scenarios, as well as challenging environmental conditions such as heavy snowfall, nighttime snow, fog, and low illumination. By providing large-scale, densely annotated video-language data under adverse winter conditions, SnowRef-Drive establishes a standardized benchmark for evaluating instruction grounding, temporal consistency, and segmentation robustness in safety-critical winter driving scenarios.