GraVoS: Voxel Selection for 3D Point-Cloud Detection

Figure 1. GraVoS for 3D object detection. Given a 3D point cloud (cyan) and its associated voxels, we propose a method that selects a subset of network-dependent meaningful voxels (salmon). Our method selects most of the voxels of the challenging classes, those with relatively few training instances (such as the Cyclists and Pedestrians), less from the prevalent classes (e.g. Cars) and very few from the background. It is shown that considering only this subset improves the performance of numerous SoTA voxel-based detectors.