Three-Dimensional Object Detection Based on Multistage Information Enhancement in Point Clouds

Shanshuai Yuan; Lei Ding

doi:10.3788/LOP223207

Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0415003(2024)

Three-Dimensional Object Detection Based on Multistage Information Enhancement in Point Clouds

Shanshuai Yuan^1,2 and Lei Ding^1,2,3、*

¹Key Laboratory of Infrared System Detection and Imaging Technology, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China

²School of Information Science and Technology, ShanghaiTech University, Shanghai 201210, China

³University of Chinese Academy of Sciences, Beijing 100049, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Overall framework of the proposed algorithm

Download full size

Fig. 2. Result after coordinate conversion when calculating IoU Loss

Download full size

Fig. 3. Feature-based network

Download full size

Fig. 4. Point-based network

Download full size

Table 1. Vehicle detection results of different algorithms on WOD validation set

View table

Table 1. Vehicle detection results of different algorithms on WOD validation set

Difficulty	Algorithm	3D AP				3D APH
Difficulty	Algorithm	Overall	Near	Middle	Far	Overall	Near	Middle	Far
L1	StarNet^［19］	55.11	80.48	48.61	27.74	54.64	79.92	48.10	27.29
	MVF^［20］	62.93	86.30	60.02	36.02
	PointPillars^［9］	63.27	84.90	59.18	35.79	62.72	84.35	58.57	35.16
	AFDet^［21］	63.69	87.38	62.19	29.27
	3D-MAN^［18］	69.03	87.99	66.55	43.15	68.52	87.57	65.92	42.37
	PV-RCNN^［11］	70.30	91.92	69.21	42.17	69.49	91.34	68.53	41.31
	CenterPoint^［10］	74.63	90.93	72.90	51.32	74.12	90.50	72.34	50.62
	Proposed algorithm	77.19	92.25	75.98	54.85	76.76	91.89	75.47	54.15
L2	StarNet^［19］	48.69	79.67	43.57	20.53	48.26	79.11	43.11	20.19
	PointPillars^［9］	55.18	83.61	53.01	26.73	54.69	83.08	52.46	26.24
	3D-MAN^［18］	60.16	87.10	59.27	32.69	59.71	86.68	58.71	32.08
	PV-RCNN^［11］	65.36	91.58	65.13	36.46	64.79	91.00	64.49	35.70
	CenterPoint^［10］	66.73	89.78	66.95	40.14	66.26	89.35	66.42	39.57
	Proposed algorithm	68.70	91.01	69.47	42.56	68.30	90.66	69.00	41.99

Table 2. Comparison between ReLU and SiLU
View table
Table 2. Comparison between ReLU and SiLU
Activation function L1 L2
3D AP 3D APH 3D AP 3D APH
ReLU 65.56 65.05 60.81 60.33
SiLU 66.21 65.71 61.41 60.95

Table 3. Comparison between the original Loss and IoU Loss
View table
Table 3. Comparison between the original Loss and IoU Loss
Note L1 L2
3D AP 3D APH 3D AP 3D APH
Original Loss 66.21 65.71 61.41 60.95
IoU Loss 66.59 66.09 61.79 61.33

Table 4. Results of different matching schemes of Feature and Point

View table

Table 4. Results of different matching schemes of Feature and Point

Scheme	L1		L2
Scheme	3D AP	3D APH	3D AP	3D APH
One stage	66.59	66.09	61.79	61.33
One stage+Feature	67.28	66.79	62.43	61.98
One stage+Point	68.60	68.16	63.61	63.19
One stage+Point+Point	68.95	68.50	63.92	63.51
One stage+Feature+Point	69.14	68.69	64.12	63.70

Table 5. Effect of using the IoU branch
View table
Table 5. Effect of using the IoU branch
Note L1 L2
3D AP 3D APH 3D AP 3D APH
No IoU Branch 69.14 68.69 64.12 63.70
With IoU Branch 69.29 68.86 64.26 63.86

Tools

Get Citation

Copy Citation Text

Shanshuai Yuan, Lei Ding. Three-Dimensional Object Detection Based on Multistage Information Enhancement in Point Clouds[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0415003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites