GeoSegNet: point cloud semantic segmentation via geometric encoder–decoder modeling

Semantic segmentation of point clouds, aiming to assign each a semantic category, is critical 3D scene understanding.Despite significant advances in recent years, most existing methods still suffer from either the object-level misclassification or boundary-level ambiguity. In this paper, we present robust network by deeply exploring geometry dubbed GeoSegNet. Our GeoSegNet consists multi-geometry based encoder and boundary-guided decoder. encoder, develop new residual module perspectives extract features. decoder, introduce contrastive boundary learning enhance geometric representation points. Benefiting encoder-decoder modeling, our can infer objects effectively while making intersections (boundaries) two more clear. Experiments show obvious improvements method over its competitors terms overall accuracy object clearness. Code available at https://github.com/Chen-yuiyui/GeoSegNet.