3DPillars: Pillar-based two-stage 3D object detection

Jongyoun Noh, Junghyup Lee, Hyekang Park, Bumsub Ham

Abstract: Highlights•We propose a novel two-stage framework for 3D object detection, dubbed 3DPillars, that extracts 3D voxel-based features efficiently using pseudo image representations without exploiting 3D convolutions.•We introduce a new RoI head with S∧<math><msup is="true"><mrow is="true"></mrow><mo is="true">∧</mo></msup></math>2CFM, enabling employing a two-stage pipeline to 3DPillars, while considering rich scene contexts to refine 3D proposals.•We demonstrate that our approach gives a better compromise in terms of speed and accuracy, compared to the state of the art.