On the Calibration of Uncertainty Estimation in LiDAR-Based Semantic Segmentation

The confidence calibration of deep learning-based perception models plays a crucial role in their reliability. Especially in the context of autonomous driving, downstream tasks like prediction and planning depend on accurate confidence estimates. In the point-wise multiclass classification task of sematic segmentation the model naturally has to deal with heavy class imbalances. Due to their underrepresentation, the confidence calibration of classes with smaller instances is challenging but essential, especially for safety reasons. We propose a metric to measure the confidence calibration quality of a semantic segmentation model with respect to individual classes. It utilizes the computation of sparsification curves based on uncertainty estimates. The insights provided by a careful analysis using the proposed metric uncover potential improvements of any semantic segmentation model and allow for a specific improvement of the model's classification and calibration performance.