Exploring Point Cloud Voxelization for 3-D Object Detection

Voxelization of point cloud data serves as a foundational process in 3D perception. Current methodologies predominantly employ uniform voxel partitioning, which proves inefficient for autonomous driving scenarios characterized by sparse point distributions. This conventional approach presents a critical trade-off dilemma: reducing spatial resolution to maintain real-time processing severely compromises small object detection accuracy, while increasing resolution for performance enhancement incurs prohibitive computational costs. To resolve this efficiency-performance paradox, we propose two innovative voxelization paradigms: 1) Density-aware <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nonuniform voxelization</b> based on spatial point distribution patterns, and 2) Context-aware <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dynamic voxelization</b> adaptively optimizes voxel allocation based on object presence or road topology. These methods enable improved performance on small objects without significantly increasing latency. Using these methods, we achieved over 2.5 APH/L2 improvement for Pedestrian/Cyclist categories on the Waymo Open Dataset, and up to 10 AP improvement for Pedestrian/Traffic Cone categories on nuScenes.