Fast Hierarchical 3D Distance Transforms on the GPU

Abstract

This paper describes a fast approximate approach for the GPU-based computation of 3D Euclidean distance transforms (DT), i.e. distance fields with associated vector information to the closest object point. Our hierarchical method works on discrete voxel grids and uses a propagation technique, both on a single hierarchy level and between the levels. Using our hierarchical approach, the effort to compute the DT is significantly reduced. It is well suited for applications that mainly rely on exact distance values close to the boundary. Our technique is purely GPU-based. All hierarchical operations are performed on the GPU. A direct comparison with the Jump Flooding Algorithm (JFA) shows that our approach is faster and provides better scaling in speed and precision, while JFA should be preferred in applications that require a more precise DT.