RoadBench

RoadBench RoadBench is a benchmark for evaluating the fine-grained spatial understanding and reasoning capabilities of multimodal large language models (MLLMs) in urban scenarios. It comprises eight tasks spanning bird’s-eye-view (BEV, satellite) and first-person-view (FPV, in-vehicle camera) imagery, including lane counting, lane designation recognition, road network correction, road type classification, and two cross-view tasks. ⚠️ Important: how BEV satellite… See the full description on the dataset page:

Source: Hugging Face Hub (tsinghua-fib-lab/RoadBench). Metadata imported from the dataset’s Hub tags.