ObsDriveBench: Benchmarking Multimodal Understanding Under Adverse Weather

ObsDriveBench has been launched by researchers as a practical multimodal benchmark aimed at assessing vision-language models for autonomous driving in challenging weather scenarios, including fog, rain, snow, and low light. This benchmark tackles the issue of diminished environmental observability, where multimodal data can become unreliable and inconsistent across different modalities. It encompasses three key dimensions: observability awareness, spatial reliability, and risk-aware decision-making, allowing for detailed analysis of model performance. The study points out that current benchmarks primarily focus on ideal conditions, synthetic distortions, or single modalities, creating a lack of insight into how models perform in real-world adverse weather with multimodal inputs. The goal is to enhance safety in autonomous driving by evaluating scene comprehension and decision-making in compromised conditions.