Range Conditioned Dilated Convolutions for Scale Invariant 3D Object\n Detection

This paper presents a novel 3D object detection framework that processes\nLiDAR data directly on its native representation: range images. Benefiting from\nthe compactness of range images, 2D convolutions can efficiently process dense\nLiDAR data of a scene. To overcome scale sensitivity in this perspective view,\na novel range-conditioned dilation (RCD) layer is proposed to dynamically\nadjust a continuous dilation rate as a function of the measured range.\nFurthermore, localized soft range gating combined with a 3D box-refinement\nstage improves robustness in occluded areas, and produces overall more accurate\nbounding box predictions. On the public large-scale Waymo Open Dataset, our\nmethod sets a new baseline for range-based 3D detection, outperforming\nmultiview and voxel-based methods over all ranges with unparalleled performance\nat long range detection.\n