Figure 1. Different sematic segmentation paradigms: (a-b) para-
metric vs (c) nonparametric. Modern segmentation solutions, no
matter using (a) parametric softmax or (b) query vectors for mask
decoding,canbeviewedaslearnableprototypebasedmethodsthat
learn class-wise prototypes in a fully parametric manner. We in-
steadproposeanonparametricscheme(c)thatdirectlyselectssub-
clustercentersofembeddedpixelsasprototypes,andachievesper-
pixel prediction via nonparametric nearest prototype retrieving.