Pointcloud to VoxelGrid: More choices and better traceability to original data

[daniel-bogdoll]

In a recent project, we use the open3d.geometry.voxelgrid to create a VoxelGrid from a pointcloud (typical lidar pointcloud from the domain of autonomous driving). The pointcloud has distinct semantic colors/labels, such as in http://www.semantic-kitti.org/.

With the "static create_from_point_cloud(input, voxel_size)" function, unfortunately two issues occur for us:

• There is no option to influence the final color choice. When there are points with different colors in a voxel, the average color is computed. This however introduces new colors, which do not have a semantic meaning, as described here: http://www.open3d.org/docs/release/python_api/open3d.geometry.VoxelGrid.html. We would love to have a max() option instead of the average() one, so that the voxels simply get the color which most of the included points have. This helps to keep the same color palette and thus the semantic meaning. Semantic VoxelGrid can look like this: https://github.com/NVlabs/VoxFormer
• There is no relation to the original pointcloud data. Often, it would be interesting to know which points were used to create a voxel.