"We propose a novel dimensionality reduction technique that uses cross-dimension and multi-scale clustering to preserve local and global structure. Combining this with representation
learning we can generate a 2D/3D latent-space that is topologically close to the real space and can be used for zone-level classification already. We demonstrate that by incorporating some real-world measurements we can transport this to a Cartesian map. As a result we are able to predict precise
2D or 3D positions for a single person in a realistic indoor environment, without using any precise position labels during training."