Dense Spatial Segmentation from Sparse Semantic Information

This paper develops an environment representation that affords reasoning about the occupancy of space, necessary for safe navigation, and about the identity of objects, necessary for complex task interpretation. The main challenge is to provide accurate dense classification of 3-D space, while observing the limitations of onboard, realtime inference and storage. Our approach constructs a graphical model of object geometry and semantics and extends this sparse graph structure into a tetrahedral decomposition of 3-D space. The resulting mesh map can be used to interpolate the sparse object properties into a dense spatial segmentation.