[Preprint] CoRE: Weakly Supervised Coarse-to-Fine Risk Evidence Learning in Driving Videos
We are pleased to share CoRE (Weakly Supervised Coarse-to-Fine Risk Evidence Learning in Driving Videos), now available as a preprint.
The work is led by Kaiser Hamid with Can Cui and Nade Liang.
CoRE studies not only how risky a driving scene is perceived to be, but also when the risk emerges and which tracked objects contribute to it. The framework learns fine-grained risk evidence using only clip-level subjective risk ratings, without requiring frame-level or object-level risk annotations.
The project transforms changes in a coarse risk predictor under structured temporal and object-level interventions into graded prediction-effect supervision for perceived-risk scoring, temporal risk-support localization, and tracked-object contribution estimation.