Abstract
Deep learning-based models have seen an exponential increase in popularity in recent years. However, they are known to be overconfident and poorly calibrated, which could result in high-confidence incorrect predictions. This is especially dangerous for high-risk applications like biomedical image processing and self-driving cars. To mitigate this, a new subfield of deep learning, uncertainty quantification (UQ), has been invented. UQ methods quantify the uncertainty associated with a model prediction, thereby allowing the identification of potentially erroneous predictions. Prior UQ work has shown promising results but often requires modifications to the model architecture or training regime, or expensive sampling at inference time. These modifications could reduce model accuracy and require model retraining, while sampling is prohibitively expensive for real-time or large-scale applications. Meta-model-based approaches have recently emerged as an effective and efficient post-hoc approach to UQ in image classification, but remain underexplored for semantic segmentation. In this work, we propose a novel meta-model-based UQ framework for image classification that mitigates several issues with prior works. We then extend it to semantic segmentation models and develop an uncertainty-aware framework for high-resolution segmentation. Experiments on two datasets show favorable results and have the potential to provide a simple, effective, and efficient way to incorporate UQ into any classification or segmentation framework.