Centroid Module for Shaping Feature Space in Semantic Segmentation

Jialei Chen, Daisuke Deguchi, Chenkai Zhang, Hiroshi Murase

Abstract: We propose Centroid Module (CM), a simple but effective module that improves the clustering ability of the image encoder in the semantic segmentation. Specifically, CM consists of a group of learnable parameters serving as the centroid of each category. During training, the features extracted by an image encoder first interact with the centroid to produce refined features. The refined features are then aligned with the corresponding centroid and the selected centroid is also fed into the classifier to be supervised by the category label. Experiments on COCO-Stuff10k datasets show impressive improvements compared with the baseline methods.