Laplace Approximation for Real-Time Uncertainty Estimation in Object Detection

Object detection is a fundamental task in autonomous driving. Besides bounding-box-like detection algorithms, uncertainty estimation is necessary for safe and trust-worthy perceptions. Bayesian Neural Networks (BNNs) provide a reliable approach to address the challenge. However, it often becomes computationally prohibitive to apply them to modern large-scale neural networks. This work develops an efficient BNN by combining the Laplace Approximation (LA) with linearized inference. Specifically, we study the effectiveness and computational necessity of a diagonal Hessian approximation in the LA on over-parameterized networks. With numerous quantitative experiments on different types of interference, the proposed method demonstrates the ability for real-time and robust uncertainty description for autonomous driving.