Abstract
When employing deep neural networks (DNNs) for semantic segmentation insafety-critical applications like automotive perception or medical imaging, itis important to estimate their performance at runtime, e.g. via uncertaintyestimates or prediction quality estimates. Previous works mostly performeduncertainty estimation on pixel-level. In a line of research, aconnected-component-wise (segment-wise) perspective was taken, approachinguncertainty estimation on an object-level by performing so-called metaclassification and regression to estimate uncertainty and prediction quality,respectively. In those works, each predicted segment is considered individuallyto estimate its uncertainty or prediction quality. However, the neighboringsegments may provide additional hints on whether a given predicted segment isof high quality, which we study in the present work. On the basis ofuncertainty indicating metrics on segment-level, we use graph neural networks(GNNs) to model the relationship of a given segment's quality as a function ofthe given segment's metrics as well as those of its neighboring segments. Wecompare different GNN architectures and achieve a notable performanceimprovement.