Compressed Voxel-Based Mapping Using Unsupervised Learning

In order to deal with the scaling problem of volumetric map representations we propose spatially local methods for high-ratio compression of 3D maps, represented as truncated signed distance fields. We show that these compres sed maps can be used as meaningful descriptors for selective decompression in scenarios relevant to robotic applicatio ns. As compression methods, we compare using PCA-derived lowdimensional bases to non-linear auto-encoder networks and novel mixed architectures that combine both. Selecting two application-oriented performance metrics, we evaluate the impact of different compression rates on reconstruction fidel ity as well as to the task of map-aided ego-motion estimation. It is demonstrated that lossily compressed distance fields use d as cost functions for ego-motion estimation, can outperfor m their uncompressed counterparts in challenging scenarios from standard RGB-D data-sets.