m2sv: A Scalable Benchmark for Map-to-Street-View Spatial Reasoning

The article introduces m2sv, a scalable benchmark for map-to-street-view spatial reasoning, consisting of m2sv-20k and m2sv-sft-11k datasets aimed at improving vision-language models (VLMs) in aligning overhead maps with Street View images. The benchmark reveals that the best VLM achieves only 65.2% accuracy, significantly lower than human annotators' average of 72.0%, indicating persistent challenges in geometric alignment and reasoning consistency. This work emphasizes the need for advancements in grounded spatial reasoning, making it crucial for practitioners focusing on multimodal AI applications.