Integrating Multiple Anomaly Assumptions: An Out-of-Distribution Detection Method in Driving Scenarios

Deep learning models struggle to recognize objects they have never encountered, often confidently classifying them as known classes. This can pose risks in certain application scenarios. A solution to this challenge is out-of-distribution (OoD) detection, which can identify objects that do not belong to known classes. However, existing methods typically rely on a single anomaly metric, leading to insufficient modeling of the original data. To address this, this paper proposes a novel anomaly metric that integrates different anomaly assumptions, enabling more reliable detection of out-of-distribution data without incurring additional computational costs. Additionally, we introduce a new loss function to optimize this metric and mitigate the issue of model overconfidence. The proposed method has been extensively tested on two driving scenario benchmark datasets, demonstrating its effectiveness.