LMSCNet: Lightweight Multiscale 3D Semantic Completion

Abstract

car road parking sidewalk building fence vegetation trunk terrain pole traffic-sign LMSCNet (1:8) (1:4) (1:2) (viz. only) Figure 1: To prevent heavy computation overhead we use a mix of 2D/3D convolutions to infer multiscale 3D semantic scene completion from sparse voxelized input. Evaluation performed on the challenging SemanticKITTI [1] benchmark shows that our LMSCNet proposal reaches state-of-the-art performance at significantly faster computation speed.