Reconstruction of 3D Semantic Map and Its Quality Estimation

Vitaly Bezuglyj, Dmitry Yudin

Abstract:

Algorithms for reconstruction of three-dimensional semantic maps are an important element of on-board vehicle computer vision systems. Such maps can be used in simulation environments and to generate the so-called HD maps needed for path planning and vehicle navigation.
The paper presents an analysis of modern methods for semantic map reconstruction based on sequences of 3D point clouds, including noisy ones. The Kimera Semantics method, modified approaches VDB Fusion, Puma and ALeGO-LOAM with Interactive SLAM are compared.
We have developed a novel approach for quantitatively estimation the quality of 3D maps and successfully applied it using the open dataset SemanticKITTI. It allows us to take into account the features of generated 3D map mesh and semantic labels in order to obtain a more informative metric.