Development of a voxelization tool for the calculation of vector-based workpiece representations

Abstract

While voxelization techniques are quite common in domains like medical imaging and computer games, they still have a lot of untapped potential when applied to CAD-models. As a discretized 3D representation, they have a significant potential for new approaches in the fields of knowledge management, workpiece classification and subsequently process planning – especially when enriched with Product Manufacturing Information (PMI). We developed a voxelization tool for colored 3D-mesh representations of workpiece geometries that allows for the structured, dimensionality-reducing storage of geometry information and PMI. The tool creates voxel-representations of triangular mesh representations by recursively splitting the space in which a mesh representation resides into eight equal cubes and checking for intersections with the mesh. If such a cube intersects with the mesh, it is again split. This process continues until a threshold cube size or recursion depth is reached. After termination, the voxelized 3D model is represented by the subset of resulting cubes containing exactly the cubes intersecting with the model’s surface. Additional functionalities have been implemented to accommodate data-centric use cases; that is, for watertight, closed meshes the tool discriminates inside and outside voxels and can include them accordingly. The calculated voxelization can also be arbitrarily sampled to a fixed voxel count – allowing to represent the 3D model by a vector of fixed length. This, for example, enables the application of AI/Machine Learning methods, the training of which typically requires a large amount of input vectors of a fixed length. Finally, we developed a plugin for the CAD application software Rhinoceros© that uses its 3D-canvas to display calculated voxelizations for visualization purposes. Going forward, data-centric applications can use workpiece information encoded inside the voxelization representation for a multitude of purposes such as similarity search or clustering.