Point Cloud 3D Object Detection Based on Multi-Scale Features and Grouped Convolutions

Fig. 6. Visualization of detection results. (a)(b) Detection performance of PV-RCNN++; (c)(d) detection performance of improved PV-RCNN++; (e)(f) 2D images

Download full size

Table 1. Performance of different algorithms on the KITTI dataset

View table

Table 1. Performance of different algorithms on the KITTI dataset

Model	Car（R_IoU=0.70）			Pedestrian（R_IoU=0.50）			Cyclist（R_IoU=0.50）			3D mAP
Model	Easy	Moderate	Hard	Easy	Moderate	Hard	Easy	Moderate	Hard	3D mAP
PointPillar	86.17	75.27	72.05	48.31	42.19	37.66	75.97	57.90	54.00	61.06
SECOND	88.54	78.73	75.86	50.78	45.37	41.08	81.63	64.24	60.10	65.15
CenterPoint	85.09	77.33	75.14	55.14	50.81	45.75	77.93	63.54	59.66	65.60
VoxelNeXt	86.38	77.13	73.10	57.14	52.24	47.21	82.51	65.77	61.51	67.00
HEDNet^［26］	87.52	79.34	77.10	58.36	52.39	48.00	80.42	62.80	58.63	67.17
SAFDNet^［27］	89.23	80.01	77.36	55.49	50.47	45.34	81.86	65.68	61.46	67.43
GCIoU	91.12	82.08	79.46	54.41	47.77	42.49	86.70	69.28	64.73	68.67
IA-SSD	87.98	79.17	76.10	56.56	52.07	47.52	88.73	70.87	66.76	69.53
PV-RCNN	91.60	82.32	79.94	59.23	52.12	47.03	85.10	68.80	64.27	70.05
Voxel-RCNN	92.32	82.34	79.81	59.88	50.94	45.68	87.15	70.97	66.83	70.66
PV-RCNN++	90.98	81.73	80.28	60.08	53.12	48.63	85.42	70.09	65.85	70.69
Proposed	91.17	82.63	80.54	63.20	57.07	52.33	89.87	73.61	68.93	73.26

Table 2. Ablation experiment results

View table

Table 2. Ablation experiment results

Method	Module	Pedestrian（R_IoU=0.50）				Cyclist（R_IoU=0.50）
Method	Module	Easy	Moderate	Hard	3D mAP	Easy	Moderate	Hard	3D mAP
Baseline	None	60.08	53.12	48.63	53.94	85.42	70.09	65.85	73.79
Experiment 1	MSF	61.41	54.81	50.44	55.55	86.18	71.81	67.87	75.29
Experiment 2	GC	61.45	55.27	50.71	55.81	89.05	72.25	68.13	76.48
Experiment 3	PKW-SimAM	61.17	54.17	49.67	55.00	87.55	71.38	67.35	75.43
Experiment 4	MSF+GC+PKW-SimAM	63.20	57.07	52.33	57.53	89.87	73.61	68.93	77.47

Table 3. Detection accuracy of the model under different segmentation methods

View table

Table 3. Detection accuracy of the model under different segmentation methods

Method	Convolution kernel					Pedestrian（3D mAP） /%	Cyclist（3D mAP） /%
Method	1×1	3×3	5×5	7×7	9×9	Pedestrian（3D mAP） /%	Cyclist（3D mAP） /%
Baseline		C	‒	‒	‒	53.94	73.79
Experiment 1		C/4	C/4	C/4	C/4	55.21	75.42
Experiment 2	C/4	C/4	C/4	C/4	‒	55.62	74.83
Experiment 3		C/2	C/4	C/8	C/8	55.42	74.96
Proposed		C/4	C/2	C/8	C/8	55.81	76.48

Table 4. Impact of different modules on the model's detection speed
View table
Table 4. Impact of different modules on the model's detection speed
Method Inference time/ms
Baseline 115.02
Baseline+MSF 121.27
Baseline+GC 131.16
Baseline+PKW-SimAM 116.70

Tools

Get Citation

Copy Citation Text

Xu Zhang, Dong Wang, Tao Wang. Point Cloud 3D Object Detection Based on Multi-Scale Features and Grouped Convolutions[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1615003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites