Point Cloud 3D Object Detection Based on Improved SECOND Algorithm

Fig. 10. Comparison of qualitative results between the proposed algorithm and the benchmark network (SECOND). (a) Visualization results of SECOND algorithm; (b) visualization results of proposed algorithm

Download full size

Fig. 11. Comparison of qualitative results of two algorithms. (a) Visualization of SECOND algorithm; (b) visualization of SECOND algorithm+ASFF

Download full size

Fig. 12. Visualization of spatial weight

Download full size

Table 1. Comparison of detection accuracy with different algorithms on KITTI test set

View table

Table 1. Comparison of detection accuracy with different algorithms on KITTI test set

Algorithm	car-3D			car-BEV			cyclist-3D			cyclist-BEV
Algorithm	easy	mod	hard	easy	mod	hard	easy	mod	hard	easy	mod	hard
Autoshape^［4］	22.47	14.17	11.36	30.66	20.08	15.95
PointRCNN^［6］	86.96	75.64	70.70	92.13	87.39	82.72	74.96	58.82	52.53	82.56	67.24	60.28
3DSSD^［7］	88.36	79.57	74.55	92.66	89.02	85.86	82.48	64.10	56.90	85.04	67.62	61.14
VoxelNet^［9］	77.47	65.11	57.73	89.35	79.26	77.39	61.22	48.36	44.37	66.70	54.76	50.55
SECOND^［10］	84.65	75.96	68.71	91.81	86.37	81.04	75.83	60.82	53.67	79.21	64.26	56.61
PointPillars^［11］	82.58	74.31	68.99	90.07	86.56	82.81	77.10	58.65	52.92	79.90	62.73	55.58
F-PointNet^［14］	82.19	69.79	60.59	91.17	84.67	74.77	72.27	56.12	49.01	77.26	61.37	53.78
PI-RCNN^［15］	84.37	74.82	70.03	91.44	85.81	81.00
AVOD-FPN^［22］	83.07	71.76	65.73	90.99	84.82	79.62	63.76	50.55	44.93	69.39	57.12	51.09
DVFENet^［23］	86.20	79.18	74.58	90.93	87.68	84.60	78.73	62.00	55.18	82.29	67.40	60.71
IA-SSD^［24］	88.34	80.13	75.04	92.79	89.33	84.35	78.35	61.94	55.70	81.30	66.29	59.58
Proposed algorithm	87.51	78.95	75.77	92.09	88.32	85.47	79.67	64.71	57.94	83.04	69.61	62.65

Table 2. Comparison of experimental results of different algorithms on the KITTI validation set

View table

Table 2. Comparison of experimental results of different algorithms on the KITTI validation set

Algorithm	Modality	FPS	car-3D
Algorithm	Modality	FPS	easy	mod	hard
F-PointNet	Lidar+RGB	9.52	83.76	70.92	63.65
SECOND	Lidar（Voxel-based）	21.74	88.16	78.18	77.04
PointRCNN	Lidar（Point-based）	9.84	88.32	78.26	77.34
PI-RCNN	Lidar（Point-based）	10.16	88.21	78.53	77.72
Proposed algorithm	Lidar（Voxel-based）	20.41	89.02	79.03	78.01

Table 3. Ablation experiments of ASFF module on the KITTI validation set

View table

Table 3. Ablation experiments of ASFF module on the KITTI validation set

Method	car-3D			car-BEV			cyclist-3D			cyclist-BEV
Method	easy	mod	hard	easy	mod	hard	easy	mod	hard	easy	mod	hard
SECOND	88.16	78.18	77.04	89.76	87.74	86.52	79.96	67.34	62.44	85.54	70.96	66.68
SECOND+ASFF	88.33	78.37	77.16	90.10	87.82	86.51	82.75	69.41	65.20	87.25	72.24	68.23
PointPillars	86.93	77.18	75.20	89.78	87.41	84.19	80.34	64.59	61.21	83.21	68.31	63.98
PointPillars+ASFF	87.80	77.74	75.23	90.15	87.75	86.04	81.90	65.09	61.53	85.86	70.80	65.69

Table 4. Ablation experiments of different loss functions on the KITTI validation set

View table

Table 4. Ablation experiments of different loss functions on the KITTI validation set

Loss	car-3D			car-BEV			cyclist-3D			cyclist-BEV
Loss	easy	mod	hard	easy	mod	hard	easy	mod	hard	easy	mod	hard
Smooth L1	88.16	78.18	77.04	89.76	87.74	86.52	79.96	67.34	62.44	85.54	70.96	66.68
3D IoU	88.51	78.60	77.51	90.02	87.99	86.70	81.91	69.02	64.22	87.89	71.92	67.79
3D GIoU	88.48	78.69	77.57	89.96	88.05	86.74	82.56	69.34	64.87	88.13	72.21	68.26
3D DIoU	88.49	78.77	77.72	89.98	88.09	86.81	83.22	69.87	65.76	88.66	72.57	68.72

Table 5. Ablation experiments of bounding box quality score on the KITTI validation set

View table

Table 5. Ablation experiments of bounding box quality score on the KITTI validation set

Score	car-3D			car-BEV			cyclist-3D			cyclist-BEV
Score	easy	mod	hard	easy	mod	hard	easy	mod	hard	easy	mod	hard
$c$	88.16	78.18	77.04	89.76	87.74	86.52	79.96	67.34	62.44	85.54	70.96	66.68
$c \cdot R_{I o U p}$	88.18	78.58	77.50	89.84	87.87	86.57	82.41	68.67	64.56	86.85	71.88	67.19

Table 6. Ablation experiments of improved SECOND algorithm on the KITTI validation set

View table

Table 6. Ablation experiments of improved SECOND algorithm on the KITTI validation set

Backbone	Loss	Score	car-3D			car-BEV			cyclist-3D			cyclist-BEV
Backbone	Loss	Score	easy	mod	hard	easy	mod	hard	easy	mod	hard	easy	mod	hard
Baseline	Smooth L1	$c$	88.16	78.18	77.04	89.76	87.74	86.52	79.96	67.34	62.44	85.54	70.96	66.68
	Smooth L1	$c \cdot R_{I o U p}$	88.18	78.58	77.50	89.84	87.87	86.57	82.41	68.67	64.56	86.85	71.88	67.19
	3D DIoU	$c$	88.49	78.77	77.72	89.98	88.09	86.81	83.22	69.87	65.76	88.66	72.57	68.72
	3D DIoU	$c \cdot R_{I o U p}$	88.77	78.89	77.98	90.11	88.03	86.94	85.35	70.77	66.88	89.77	73.46	69.94
ASFF	Smooth L1	$c$	88.33	78.37	77.16	90.10	87.82	86.51	82.75	69.41	65.20	87.25	72.24	68.23
	Smooth L1	$c \cdot R_{I o U p}$	88.47	78.65	77.60	89.91	87.95	86.66	84.36	70.11	66.08	88.92	72.82	68.89
	3D DIoU	$c$	88.67	78.81	77.74	90.06	88.17	86.87	84.82	70.26	66.35	89.34	73.15	69.28
	3D DIoU	$c \cdot R_{I o U p}$	89.02	79.03	78.01	90.20	88.27	87.26	86.45	71.75	67.61	91.32	74.12	70.77

Tools

Get Citation

Copy Citation Text

Ying Zhang, Liangliang Jiang, Dongbo Zhang, Wanlin Duan, Yue Sun. Point Cloud 3D Object Detection Based on Improved SECOND Algorithm[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0811005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Imaging Systems

Received: Apr. 3, 2023

Accepted: Aug. 2, 2023

Published Online: Mar. 15, 2024

The Author Email: Jiang Liangliang (1017204759@qq.com)

DOI:10.3788/LOP231016

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology