CertainNet: Sampling-Free Uncertainty Estimation for Object Detection

Estimating the uncertainty of a neural network plays fundamental role in safety-critical settings. In perception for autonomous driving, measuring means providing additional calibrated information to downstream tasks, such as path planning, that can use it towards safe navigation. this work, we propose novel sampling-free estimation method object detection. We call CertainNet, and is first provide separate uncertainties each output signal: objectness, class, location size. To achieve this, an uncertainty-aware heatmap, exploit neighboring bounding boxes provided by detector at inference time. evaluate detection performance quality different estimates separately, also with challenging out-of-domain samples: BDD100K nuImages models trained on KITTI. Additionally, new metric size uncertainties. When transferring unseen datasets, CertainNet generalizes substantially better than previous methods ensemble, while being real-time high comprehensive estimates.