Roldão Luis ; de Charette Raoul ; Verroust-Blondet Anne

Abstract

We introduce a new approach for multiscale 3Dsemantic scene completion fromvoxelized sparse 3D LiDAR scans. As opposed to the literature, we use a 2D UNetbackbone with comprehensive multiscale skip connections to enhance featureflow, along with 3D segmentation heads. On the SemanticKITTI benchmark, ourmethod performs on par on semantic completion and better on occupancycompletion than all other published methods -- while being significantlylighter and faster. As such it provides a great performance/speed trade-off formobile-robotics applications. The ablation studies demonstrate our method isrobust to lower density inputs, and that it enables very high speed semanticcompletion at the coarsest level. Our code is available athttps://github.com/cv-rits/LMSCNet.