BD-HazardVLM: Probing Vision-Language Models for Latent Defensive-Driving Hazards in Bangladesh Road Scenes
It is shown that coarse hazard recognition does not imply reliable fine-grained defensive-driving reasoning and motivate geographically diverse safety evaluation, and all models show weak latent-hazard anticipation and near-zero self-reported trap recognition.
A. Al Maruf, Md. Sajedul Islam, Tanmoy Mridha et al.
· 0 citations