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

Table 3

Detection performance comparison among several methods on KITTI validation set (car class).

Methods	Modality	Anchor free	Stage	Speed (Hz)	3D detection AP ()			BEV detection AP ()
					Easy	Moderate	Hard	Easy	Moderate	Hard
F-PointNet [20]	L+C	N	Two	5.9	83.76	70.92	63.65	88.16	84.02	76.44
PointRCNN [21]	L	N	Two	10	89.19	78.85	77.91	90.21	87.89	85.51
Fast-PointRCNN [22]	L	N	Two	15.4	89.12	79	77.48	90.12	88.1	86.24
VoxelNet [12]	L	N	One	4.4	81.97	65.46	62.85	89.6	84.81	78.57
SECOND [13]	L	N	One	20	87.43	76.48	69.1	89.96	87.07	79.66
PointPillars [14]	L	N	One	42	87.44	77.67	75.76	89.88	87.43	85.01
AFDet [17]	L	Y	One	35	85.68	75.57	69.31	89.42	85.45	80.56
Ours	L	Y	One	35.7	88.65	78.20	76.17	90.04	87.80	84.14