EFNet: enhancing feature information for 3D object detection in LiDAR point clouds

Xin Meng; Yuan Zhou; Kaiyue Du; Jun Ma; Jin Meng; Aakash Kumar; Jiahang Lv; Jonghyuk Kim; Shifeng Wang; Shifeng Wang

doi:10.1364/JOSAA.511948

Journal of the Optical Society of America A
Vol. 41,
Issue 4,
pp. 739-748
(2024)
•https://doi.org/10.1364/JOSAA.511948

EFNet: enhancing feature information for 3D object detection in LiDAR point clouds

Xin Meng, Yuan Zhou, Kaiyue Du, Jun Ma, Jin Meng, Aakash Kumar, Jiahang Lv, Jonghyuk Kim, and Shifeng Wang

Not Accessible

Your library or personal account may give you access

Get PDF
Email
Share
Get Citation
Copy Citation Text
Xin Meng, Yuan Zhou, Kaiyue Du, Jun Ma, Jin Meng, Aakash Kumar, Jiahang Lv, Jonghyuk Kim, and Shifeng Wang, "EFNet: enhancing feature information for 3D object detection in LiDAR point clouds," J. Opt. Soc. Am. A 41, 739-748 (2024)

Export Citation
- BibTex
- Endnote (RIS)
- HTML
- Plain Text
Citation alert
Save article

Check for updates

Abstract

With the development of autonomous driving, there has been considerable attention on 3D object detection using LiDAR. Pillar-based LiDAR point cloud detection algorithms are extensively employed in the industry due to their simple structure and high real-time performance. Nevertheless, the pillar-based detection network suffers from significant loss of 3D coordinate information during the feature degradation and extraction process. In the paper, we introduce a novel framework with high performance, termed EFNet. The EFNet uses the Enhancing Pillar Feature Module (EPFM) to provide more accurate representations of features from two directions: pillar internal space and pillar external space. Additionally, the Head Up Module (HUM) is utilized in the detection head to integrate multi-scale information and enhance the network’s information perception ability. The EFNet achieves impressive results on the nuScenes datasets, namely, 53.3% NDS and 42.4% mAP. Compared to the baseline PointPillars, EFNet improves 8% NDS and 11.9% mAP. The results demonstrate that the proposed framework can effectively improve the network’s accuracy while ensuring deployability.

Full Article | PDF Article

More Like This

LiDAR-camera-system-based unsupervised and weakly supervised 3D object detection

Haosen Wang, Tiankai Chen, Xiaohang Ji, Feng Qian, Yue Ma, and Shifeng Wang
J. Opt. Soc. Am. A 40(10) 1849-1860 (2023)

PointCartesian-Net: enhancing 3D coordinates for semantic segmentation of large-scale point clouds

Yuan Zhou, Qi Sun, Jin Meng, Qinglong Hu, Jiahang Lyu, Zhiwei Wang, and Shifeng Wang
J. Opt. Soc. Am. A 38(8) 1194-1200 (2021)

UMA-Net: an unsupervised representation learning network for 3D point cloud classification

Jie Liu, Yu Tian, Guohua Geng, Haolin Wang, Da Song, Kang Li, Mingquan Zhou, and Xin Cao
J. Opt. Soc. Am. A 39(6) 1085-1094 (2022)

Data availability

Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the author upon reasonable request.

Cited By

You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Figures (5)

You do not have subscription access to this journal. Figure files are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Tables (6)

You do not have subscription access to this journal. Article tables are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Equations (6)

You do not have subscription access to this journal. Equations are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

		${A P}_{3 D}$				${A P}_{B E V}$
Methods	Class	Easy	Mod	Hard	mAP	Easy	Mod	Hard	mAP
VoxelNet [7]	Car	81.79	65.46	62.85	70.09	89.60	84.81	78.57	84.33
PointPillars [2]		86.99	77.12	74.98	79.70	89.77	87.02	83.71	86.83
EFNet		86.71	77.25	75.73	79.90	89.66	87.27	85.61	87.51
VoxelNet [7]	Ped.	57.86	53.42	48.87	53.38	65.95	61.05	56.98	61.33
PointPillars [2]		52.94	48.99	45.05	48.99	57.77	53.36	49.89	53.67
EFNet		58.31	53.66	48.71	53.56	63.11	56.94	53.67	57.91
VoxelNet [7]	Cyc.	67.17	47.65	45.11	53.31	74.41	52.18	50.49	59.03
PointPillars [2]		78.77	62.70	58.95	66.81	80.89	66.50	62.10	69.83
EFNet		79.76	63.68	60.58	68.01	83.02	67.61	62.51	71.05

Methods	NDS (%)	mAP (%)	Car	Tru.	Bus	Tra.	C.V.	Ped.	Mot.	Byc.	T.C.	Bar.
BirdNet $+$ [27]	29.6	24.5	44.5	29.5	31.3	31.5	10.4	31.2	21.7	3.3	3.0	38.7
InfoFocus [28]	39.5	39.5	77.9	31.4	44.8	37.3	10.7	63.4	29.0	6.1	46.5	47.8
WYSIWYG [29]	41.9	35.0	79.1	30.4	46.6	40.1	7.1	65.0	18.2	0.1	28.8	34.7
PointPillars [2]	45.3	30.5	68.4	23.0	28.2	23.4	4.1	59.7	27.4	1.1	30.8	38.9
SARPNET [30]	48.4	32.4	59.9	18.7	19.4	18.0	11.6	69.4	29.8	14.2	44.6	38.3
Reconfig V [31]	50.6	32.5	75.8	27.2	39.5	38.0	6.5	62.5	15.2	0.2	25.7	34.9
EFNet (S)	53.3	42.4	80.9	44.6	54.6	46.0	10.4	68.6	28.7	5.0	42.7	41.9
EFNet (T)	62.0	52.8	82.9	46.7	57.6	50.2	14.3	77.0	45.0	20.5	67.1	66.7

Methods	NDS (%)	mAP (%)	Car	Tru.	Bus	Tra.	C.V.	Ped.	Mot.	Byc.	T.C.	Bar.
PointPillars (S)	51.3	37.5	78.0	39.2	56.2	32.0	10.8	62.4	25.6	1.7	30.5	38.7
EFNet (S)	52.3	39.1	78.7	41.6	57.9	31.7	9.4	66.4	28.1	3.0	34.0	39.9
PointPillars (M)	56.9	42.5	81.4	49.1	62.3	33.1	11.7	70.1	31.8	3.1	36.5	45.7
EFNet (M)	58.0	44.6	81.1	49.9	61.4	33.5	11.9	72.4	33.5	19.1	47.7	35.2
PointPillars (C)	59.6	48.9	83.2	50.5	62.4	33.0	11	77.7	43.7	17.5	52.8	58.1
EFNet (C)	60.5	50.3	83.3	51.6	63.2	29.0	11.9	79.1	46.5	22.8	56.5	59.4
PointPillars (T)	60.8	50.3	83.6	52.9	63.0	34.6	13.5	76.5	44.8	18.4	50.6	64.7
EFNet (T)	62.6	52.8	84.6	53.7	66.5	36.0	13.7	80.1	49.8	25.8	55.8	64.4

	$m A P_{3 D}$			$m A P_{B E V}$
Methods	Car	Ped.	Cyc.	Car	Ped.	Cyc.
Baseline	79.70	48.99	66.81	86.83	53.67	69.83
HUM	79.84 ( $+$ 0.14)	52.70 ( $+$ 3.71)	68.58 ( $+$ 1.77)	87.44 ( $+$ 0.61)	57.31 ( $+$ 3.64)	72.36 ( $+$ 2.53)
EPFM	79.69 (−0.01)	51.33 ( $+$ 2.34)	69.50 ( $+$ 2.69)	87.19 ( $+$ 0.36)	56.90 ( $+$ 3.23)	72.49 ( $+$ 2.66)
EFNet	79.90 ( $+$ 0.20)	53.56 ( $+$ 4.57)	68.01 ( $+$ 1.20)	87.51 ( $+$ 0.68)	57.91 ( $+$ 4.24)	71.05 ( $+$ 1.22)

Methods	NDS (%)	mAP (%)	Car	Tru.	Bus	Tra.	C.V.	Ped.	Mot.	Byc.	T.C.	Bar.
Baseline	51.3	37.5	78.0	39.2	56.2	32.0	10.8	62.4	25.6	1.7	30.5	38.7
HUM	51.5 ( $+$ 0.2)	38.5 ( $+$ 1.0)	78.4 ( $+$ 0.4)	39.5 ( $+$ 0.3)	57.5 ( $+$ 1.3)	33.5 ( $+$ 1.5)	11.4 ( $+$ 0.6)	63.3 ( $+$ 0.9)	26.7 ( $+$ 1.1)	2.8 ( $+$ 1.1)	31.1 ( $+$ 0.6)	41.3 ( $+$ 2.6)
EPFM	52.0 ( $+$ 0.7)	38.4 ( $+$ 0.9)	78.6 ( $+$ 0.6)	39.2 ( $+$ 0.0)	59.2 ( $+$ 3.0)	31.2 (−0.8)	9.4 (−1.4)	64.8 ( $+$ 2.4)	25.6 ( $+$ 0.0)	2.6 ( $+$ 0.9)	33.6 ( $+$ 3.1)	39.8 ( $+$ 1.1)
EFNet	52.3 ( $+$ 1.0)	39.1 ( $+$ 1.6)	78.7 ( $+$ 0.7)	41.6 ( $+$ 2.4)	57.9 ( $+$ 1.7)	31.7 (−0.3)	9.4 (−1.4)	66.4 ( $+$ 4.0)	28.1 ( $+$ 2.5)	3.0 ( $+$ 1.3)	34.0 ( $+$ 3.5)	39.9 ( $+$ 1.2)

	Car			Ped.			Cyc.
Methods	Easy	Mod.	Hard	Easy	Mod.	Hard	Easy	Mod.	Hard	mAP
EFNet [3,5,5]	86.71	77.25	75.53	58.31	53.66	48.71	79.76	63.68	60.58	67.17
EFNet [6,10,10]	85.62	76.75	75.41	61.74	55.73	51.25	81.39	64.36	61.19	68.16
EFNet (0.5, 0.35)	86.71	77.25	75.53	58.31	53.66	48.71	79.76	63.68	60.58	67.17
EFNet (0.4, 0.25)	85.65	76.53	74.55	61.35	55.87	50.34	77.06	61.58	58.21	66.79

Abstract

Data availability

Cited By

Figures (5)

Tables (6)

Equations (6)

Journal of the Optical Society of America A