Semantic segmentation via pixel‐to‐center similarity calculation

Since the fully convolutional network has achieved great success in semantic segmentation, lots of works have been proposed to extract discriminative pixel representations. However, authors observe that existing methods still suffer from two typical challenges: (i) The intra-class feature variation between different scenes may be large, leading difficulty maintaining consistency same-class pixels scenes; (ii) inter-class distinction same scene could small, resulting limited performance distinguish classes each scene. first rethink segmentation a perspective similarity and class centers. Each weight vector head represents its corresponding whole dataset, which can regarded as embedding center. Thus, pixel-wise classification amounts computing final space Under this novel view, propose Class Center Similarity (CCS) layer address above-mentioned challenges by generating adaptive centers conditioned on supervising similarities CCS utilises Adaptive Module generate scene, adapt large scenes. Specially designed Distance Loss (CD Loss) is introduced control both distances based predicted center-to-center pixel-to-center similarity. Finally, outputs processed prediction. Extensive experiments demonstrate our model performs favourably against state-of-the-art methods.