Sparse Volumes for Large Scale 3D Modelling

Modern emergence of automation in the industry and everyday life is leveraged by extensive research in mobile robotics. Novel 3D sensors such as laser scanners or cameras enable cars to drive autonomously, UAVs to observe critical environments, or an underwater robot to construct pipelines. However, 3D sensor samples do not provide the intrinsic information a robot needs to operate on. Voxel based shape modelling has been identified as a fruitful solution. However, its application is limited to small areas since processing and visualization of large environments is very challenging. Dense voxel grids allow fast data access but suffer from a large memory overhead. Modelling an area of 100x100x100m with a resolution of 1cm would result in a 3.7TB memory requirement. Motivated by this, sparse voxel octrees (SVO) [4] have been proposed. These however, increase the data access complexity fro