An Anchor-Free 3D Object Detection Approach Based on Hierarchical Pillars

Table 1

Summary of methods for 3D object detection.

Methods	Data representation	Detector	Key innovation
[12]	Voxel-based	Anchor-based	(1) Voxel feature encoding layer; (2) 3D CNN
[13]	Voxel-based	Anchor-based	(1) 3D sparse convolution; (2) Better data augmentation
[14]	Pillar-based	Anchor-based	(1) Voxel division without vertical direction; (2) Avoid 3D convolution operation
[25]	Pillar-based	Anchor-free	(1) Aligned pillar-to-point projection; (2) Multiview feature learning
[17]	Voxel-based	Anchor-free	Anchor-free detector
[26]	Point-based	Anchor-based	Deep Hough voting
[20]	Point-based	Anchor-based	(1) Introduction of RGB images; (2) Frustum projection
[21]	Point-based	Anchor-based	(1) Rough 3D proposals based on raw point cloud; (2) Fuse local spatial features and global semantic features on the second stage
[22]	Voxel-based	Anchor-based	(1) Voxel-based first-stage prediction; (2) Combined with point features for refinement
[23]	Point-based	Anchor-based	(1) Spherical anchor; (2) PointsPool
[24]	Point-based	Anchor-free	(3) Feature-based furthest point sampling (F-FPS); (4) 3D centerness assignment strategy