||Finding the right scales for feature extraction is crucial for pixel-based supervised image segmentation. Lindeberg's scale selection scheme and its variations are applied widely within the computer vision and image analysis community, but they have been developed as unsupervised techniques. We consider the problem of selecting scales in image segmentation and aim to account for the fact that we rely on supervision, i.e., want to discriminate between predefined classes. We show the deficiency of the classical unsupervised scale selection paradigm and present a supervised alternative. Our proposal is mainly conceptual in nature and we study simple segmentation settings. In particular, we propose the so-called max rule, which selects that scale for each pixel that has the largest confidence in classification across scale. Interpreting the classifier as a complex image filter, we relate our approach back to Lindeberg's original proposal.