Pointpillars point cloud detection network based on knowledge distillation and location guidance

Jing ZHAO; Shaobo LI; Jielong GUO; Hui YU; Jianfeng ZHANG; Jie LI

doi:10.37188/CJLCD.2023-0058

Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 1, 79(2024)

Pointpillars point cloud detection network based on knowledge distillation and location guidance

Jing ZHAO^1,4, Shaobo LI^1,2, Jielong GUO^2,3、*, Hui YU^2,3, Jianfeng ZHANG^2,3, and Jie LI^2,3

Author Affiliations

¹School of Electrical Engineering and Automation，Xiamen University of Technology，Xiamen 361024，China

²Fujian Institute of Research on the Structure of Matter，Chinese Academy of Sciences，Fuzhou 350108，China

³Quanzhou Institute of Equipment Manufacturing，Haixi Institutes，Chinese Academy of Sciences，Quanzhou 362000，China

⁴Xiamen Key Laboratory of Frontier Electric Power Equipment and Intelligent Controly，Xiamen 361024，China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Network block diagram

Download full size

View in Article

Fig. 2. Positioning guidance classification

Download full size

View in Article

Fig. 3. Comparison of network model reasoning speed

Download full size

View in Article

Fig. 4. Comparison of average precision of car class

Download full size

View in Article

Fig. 5. Comparison of average accuracy of three difficulty levels

Download full size

View in Article

Table 1. Comparison of 3D detection accuracy（3D_R40）of different algorithms in KITTI dataset

View table

View in Article

Table 1. Comparison of 3D detection accuracy（3D_R40）of different algorithms in KITTI dataset

	算法	数据类型	Car（IoU=0.7）			Pedestrian（IoU=0.5）			Cyclist（IoU=0.5）
	算法	数据类型	简单	中等	困难	简单	中等	困难	简单	中等	困难
两阶段	AVOD^［19］	L+R	83.07	71.76	65.73	36.10	27.86	25.76	57.19	42.08	38.29
	PointRCNN^［5］	L	86.96	75.64	70.70	49.43	41.78	38.63	73.93	59.60	53.59
	UberATG-MMF^［20］	L+R	88.40	77.43	70.22	N/A	N/A	N/A	N/A	N/A	N/A
	Part-A2^［21］	L	87.81	78.49	73.51	53.10	43.35	40.06	79.17	63.52	56.93
单阶段	SECOND^［18］	L	83.34	72.55	65.82	51.07	42.56	37.29	70.51	53.85	46.90
	TANet^［22］	L	84.39	75.94	68.82	53.72	44.34	40.49	75.70	59.44	52.53
	Associate-3Det^［23］	L	85.99	77.40	70.53	N/A	N/A	N/A	N/A	N/A	N/A
	Point-GNN^［24］	L	88.33	78.47	72.29	51.92	43.77	40.14	78.60	63.48	57.08
	Ours	L	88.15	78.95	74.97	52.77	46.09	41.09	81.66	61.31	57.21

Table 2. Comparison of BEV detection accuracy（BEV_R40）of different algorithms in KITTI dataset

View table

View in Article

Table 2. Comparison of BEV detection accuracy（BEV_R40）of different algorithms in KITTI dataset

	算法	数据类型	Car（IoU=0.7）			Pedestrian（IoU=0.5）			Cyclist（IoU=0.5）
	算法	数据类型	简单	中等	困难	简单	中等	困难	简单	中等	困难
两阶段	AVOD^［19］	L+R	89.75	84.95	78.32	42.58	33.57	30.14	64.11	48.15	42.37
	PointRCNN^［5］	L	92.13	87.39	82.72	N/A	N/A	N/A	N/A	N/A	N/A
	UberATG-MMF^［20］	L+R	93.67	88.21	81.99	N/A	N/A	N/A	N/A	N/A	N/A
	Part-A2^［21］	L	91.70	87.79	84.41	59.04	49.81	45.92	83.43	68.73	61.85
单阶段	SECOND^［18］	L	89.39	83.77	78.59	55.10	46.27	44.76	73.67	56.04	48.78
	TANet^［22］	L	91.58	86.54	81.19	60.85	51.38	47.54	79.16	63.77	56.21
	Associate-3Det^［23］	L	91.40	88.09	82.96	N/A	N/A	N/A	N/A	N/A	N/A
	Point-GNN^［24］	L	93.11	89.17	83.90	55.36	47.07	44.61	81.17	67.28	59.67
	Ours	L	93.09	88.86	84.50	58.46	51.88	47.43	84.10	64.03	59.82

Table 3. Influence of temperature coefficient on model detection accuracy in distillation under Mod_R40
View table
View in Article
Table 3. Influence of temperature coefficient on model detection accuracy in distillation under Mod_R40
$τ$ Car Pedestrian Cyclist
教师网络 74.31 41.92 51.92
2 78.95 56.09 61.31
3 76.33 44.56 59.89
4 75.22 43.67 58.02
5 73.87 41.03 46.63

Table 4. Ablation experiments of regression frame distillation and location-guided classification in KITTI dataset

View table

View in Article

Table 4. Ablation experiments of regression frame distillation and location-guided classification in KITTI dataset

	PGC	RBD	Car（IoU=0.7）				Pedestrian（IoU=0.5）				Cyclist（IoU=0.5）
	PGC	RBD	简单	中等	困难	mAP	简单	中等	困难	mAP	简单	中等	困难	mAP
3D	×	×	82.58	74.31	68.99	75.29	51.45	41.92	38.89	44.08	77.10	58.65	51.92	62.55
	√	×	85.04	76.29	72.04	77.79	51.50	44.11	39.65	45.08	77.60	60.30	53.22	63.71
	×	√	88.08	77.10	74.12	79.77	52.28	45.79	40.97	46.35	77.58	59.17	55.37	64.04
	√	√	88.15	78.95	74.97	80.69	52.77	46.09	41.09	46.65	81.66	61.31	57.21	66.73
BEV	×	×	90.07	86.56	82.81	86.48	57.60	48.64	45.78	50.67	79.90	62.73	55.58	66.07
	√	×	91.37	87.52	82.95	88.28	57.64	50.73	46.27	51.45	82.05	63.76	60.62	70.81
	×	√	92.51	88.51	83.93	88.32	58.89	52.67	47.85	53.14	83.33	64.29	60.51	69.38
	√	√	93.09	88.86	84.50	88.82	58.46	51.88	47.43	52.59	84.10	64.03	59.82	69.32

Tools

Get Citation

Copy Citation Text

Jing ZHAO, Shaobo LI, Jielong GUO, Hui YU, Jianfeng ZHANG, Jie LI. Pointpillars point cloud detection network based on knowledge distillation and location guidance[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(1): 79

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Feb. 17, 2023

Accepted: --

Published Online: Mar. 27, 2024

The Author Email: Jielong GUO (gjl@fjirsm.ac.cn)

DOI:10.37188/CJLCD.2023-0058

Topics

Table 1. Comparison of 3D detection accuracy（3DR40）of different algorithms in KITTI dataset

Table 1. Comparison of 3D detection accuracy（3DR40）of different algorithms in KITTI dataset

Table 2. Comparison of BEV detection accuracy（BEVR40）of different algorithms in KITTI dataset

Table 2. Comparison of BEV detection accuracy（BEVR40）of different algorithms in KITTI dataset

Table 3. Influence of temperature coefficient on model detection accuracy in distillation under ModR40

Table 3. Influence of temperature coefficient on model detection accuracy in distillation under ModR40

Table 4. Ablation experiments of regression frame distillation and location-guided classification in KITTI dataset

Table 4. Ablation experiments of regression frame distillation and location-guided classification in KITTI dataset

Table 1. Comparison of 3D detection accuracy（3D_R40）of different algorithms in KITTI dataset

Table 1. Comparison of 3D detection accuracy（3D_R40）of different algorithms in KITTI dataset

Table 2. Comparison of BEV detection accuracy（BEV_R40）of different algorithms in KITTI dataset

Table 2. Comparison of BEV detection accuracy（BEV_R40）of different algorithms in KITTI dataset

Table 3. Influence of temperature coefficient on model detection accuracy in distillation under Mod_R40

Table 3. Influence of temperature coefficient on model detection accuracy in distillation under Mod_R40