A Review of Segmentation Methods for 3D Semantic Mapping

A 3D semantic map can be defined as a grid-based representation of the environment, where each bin stores a probability distribution over the possible elements to be found in it. This probability distribution can be obtained with any state-of-the-art image classifier, while the 3D position depends on the localization accuracy of the robot, the sensitivity of its RGB-D sensor, and the segmentation of the input image. In this paper, we focus on this last factor, to explore different options for image segmentation that might improve 3D maps. We will compare various approaches based on the use of 2D and 3D information to find relevant clusters of information. They will be evaluated to assess their suitability for real-time applications.