RiskBench: A Scenario-based Benchmark for Risk Identification
ImageBounding Box
Within RiskBench, I worked on detailed spatio‑temporal risk annotations. For each scenario, we labeled risk objects with precise bounding boxes and marked hazardous areas with polygons (e.g., conflict zones, potential collision regions) on a per‑frame basis over time. This required identifying not only which objects are risky, but also exactly when they become dangerous in the sequence, so all annotations were done frame‑by‑frame and aligned with the underlying trajectories. These fine‑grained labels enable models to learn both object‑level risk (who is dangerous and when) and region‑level risk (where it is unsafe to drive) in dynamic traffic scenes.
Within RiskBench, I worked on detailed spatio‑temporal risk annotations. For each scenario, we labeled risk objects with precise bounding boxes and marked hazardous areas with polygons (e.g., conflict zones, potential collision regions) on a per‑frame basis over time. This required identifying not only which objects are risky, but also exactly when they become dangerous in the sequence, so all annotations were done frame‑by‑frame and aligned with the underlying trajectories. These fine‑grained labels enable models to learn both object‑level risk (who is dangerous and when) and region‑level risk (where it is unsafe to drive) in dynamic traffic scenes.
2021 - 2023