BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection

Yifan Wu; Hongwen He; Yingjuan Tang

doi:10.1007/978-981-95-4875-0_13

BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection

Yifan Wu
, Hongwen He^*
, Yingjuan Tang

^*Corresponding author for this work

Beijing Institute of Technology

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Fusing LiDAR and camera data for 3D object detection remains a key challenge in autonomous driving. While most methods adopt dual-branch frameworks to extract BEV features from both modalities before fusion, progress in point cloud feature extraction still lags behind that of image-based networks, limiting overall fusion effectiveness. To address this gap, we propose BEV-SI, a novel multi-modal detection framework featuring a lightweight yet expressive LiDAR branch. At its core is the Split-Inception Block, which enhances point cloud representation by applying diverse channel-wise operations and expanding the receptive field. Furthermore, we introduce the Split-Neck module, which performs efficient multi-scale feature fusion through adaptive downsampling and Branch Attention, allowing the network to dynamically reweight spatial features across different scales. Extensive experiments on the nuScenes benchmark demonstrate that BEV-SI achieves competitive accuracy with significantly improved inference speed.

Original language	English
Title of host publication	Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers
Editors	Huiyun Li, Zhongli Wang, Shuai Zhao, Peng Sun, Michael Herrmann, Xi Zheng, Yuling Liu
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	160-171
Number of pages	12
ISBN (Print)	9789819548743
DOIs	https://doi.org/10.1007/978-981-95-4875-0_13
Publication status	Published - 2026
Externally published	Yes
Event	3rd CCF Intelligent Vehicles Symposium, CIVS 2025 - Hangzhou, China Duration: 16 Aug 2025 → 18 Aug 2025

Publication series

Name	Communications in Computer and Information Science
Volume	2631 CCIS
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	3rd CCF Intelligent Vehicles Symposium, CIVS 2025
Country/Territory	China
City	Hangzhou
Period	16/08/25 → 18/08/25

Keywords

3D object detection
Autonomous driving
Multi-modal fusion
Multi-scale fusion

Access to Document

10.1007/978-981-95-4875-0_13

Cite this

Wu, Y., He, H., & Tang, Y. (2026). BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection. In H. Li, Z. Wang, S. Zhao, P. Sun, M. Herrmann, X. Zheng, & Y. Liu (Eds.), Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers (pp. 160-171). (Communications in Computer and Information Science; Vol. 2631 CCIS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-95-4875-0_13

Wu, Yifan ; He, Hongwen ; Tang, Yingjuan. / BEV-SI : A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection. Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers. editor / Huiyun Li ; Zhongli Wang ; Shuai Zhao ; Peng Sun ; Michael Herrmann ; Xi Zheng ; Yuling Liu. Springer Science and Business Media Deutschland GmbH, 2026. pp. 160-171 (Communications in Computer and Information Science).

@inproceedings{7259112e19d64949a60d0d128545e7d8,

title = "BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection",

abstract = "Fusing LiDAR and camera data for 3D object detection remains a key challenge in autonomous driving. While most methods adopt dual-branch frameworks to extract BEV features from both modalities before fusion, progress in point cloud feature extraction still lags behind that of image-based networks, limiting overall fusion effectiveness. To address this gap, we propose BEV-SI, a novel multi-modal detection framework featuring a lightweight yet expressive LiDAR branch. At its core is the Split-Inception Block, which enhances point cloud representation by applying diverse channel-wise operations and expanding the receptive field. Furthermore, we introduce the Split-Neck module, which performs efficient multi-scale feature fusion through adaptive downsampling and Branch Attention, allowing the network to dynamically reweight spatial features across different scales. Extensive experiments on the nuScenes benchmark demonstrate that BEV-SI achieves competitive accuracy with significantly improved inference speed.",

keywords = "3D object detection, Autonomous driving, Multi-modal fusion, Multi-scale fusion",

author = "Yifan Wu and Hongwen He and Yingjuan Tang",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.; 3rd CCF Intelligent Vehicles Symposium, CIVS 2025 ; Conference date: 16-08-2025 Through 18-08-2025",

year = "2026",

doi = "10.1007/978-981-95-4875-0\_13",

language = "English",

isbn = "9789819548743",

series = "Communications in Computer and Information Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "160--171",

editor = "Huiyun Li and Zhongli Wang and Shuai Zhao and Peng Sun and Michael Herrmann and Xi Zheng and Yuling Liu",

booktitle = "Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers",

address = "Germany",

}

Wu, Y, He, H & Tang, Y 2026, BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection. in H Li, Z Wang, S Zhao, P Sun, M Herrmann, X Zheng & Y Liu (eds), Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers. Communications in Computer and Information Science, vol. 2631 CCIS, Springer Science and Business Media Deutschland GmbH, pp. 160-171, 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Hangzhou, China, 16/08/25. https://doi.org/10.1007/978-981-95-4875-0_13

BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection. / Wu, Yifan; He, Hongwen; Tang, Yingjuan.
Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers. ed. / Huiyun Li; Zhongli Wang; Shuai Zhao; Peng Sun; Michael Herrmann; Xi Zheng; Yuling Liu. Springer Science and Business Media Deutschland GmbH, 2026. p. 160-171 (Communications in Computer and Information Science; Vol. 2631 CCIS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - BEV-SI

T2 - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025

AU - Wu, Yifan

AU - He, Hongwen

AU - Tang, Yingjuan

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

PY - 2026

Y1 - 2026

N2 - Fusing LiDAR and camera data for 3D object detection remains a key challenge in autonomous driving. While most methods adopt dual-branch frameworks to extract BEV features from both modalities before fusion, progress in point cloud feature extraction still lags behind that of image-based networks, limiting overall fusion effectiveness. To address this gap, we propose BEV-SI, a novel multi-modal detection framework featuring a lightweight yet expressive LiDAR branch. At its core is the Split-Inception Block, which enhances point cloud representation by applying diverse channel-wise operations and expanding the receptive field. Furthermore, we introduce the Split-Neck module, which performs efficient multi-scale feature fusion through adaptive downsampling and Branch Attention, allowing the network to dynamically reweight spatial features across different scales. Extensive experiments on the nuScenes benchmark demonstrate that BEV-SI achieves competitive accuracy with significantly improved inference speed.

AB - Fusing LiDAR and camera data for 3D object detection remains a key challenge in autonomous driving. While most methods adopt dual-branch frameworks to extract BEV features from both modalities before fusion, progress in point cloud feature extraction still lags behind that of image-based networks, limiting overall fusion effectiveness. To address this gap, we propose BEV-SI, a novel multi-modal detection framework featuring a lightweight yet expressive LiDAR branch. At its core is the Split-Inception Block, which enhances point cloud representation by applying diverse channel-wise operations and expanding the receptive field. Furthermore, we introduce the Split-Neck module, which performs efficient multi-scale feature fusion through adaptive downsampling and Branch Attention, allowing the network to dynamically reweight spatial features across different scales. Extensive experiments on the nuScenes benchmark demonstrate that BEV-SI achieves competitive accuracy with significantly improved inference speed.

KW - 3D object detection

KW - Autonomous driving

KW - Multi-modal fusion

KW - Multi-scale fusion

UR - https://www.scopus.com/pages/publications/105028087249

U2 - 10.1007/978-981-95-4875-0_13

DO - 10.1007/978-981-95-4875-0_13

M3 - Conference contribution

AN - SCOPUS:105028087249

SN - 9789819548743

T3 - Communications in Computer and Information Science

SP - 160

EP - 171

BT - Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers

A2 - Li, Huiyun

A2 - Wang, Zhongli

A2 - Zhao, Shuai

A2 - Sun, Peng

A2 - Herrmann, Michael

A2 - Zheng, Xi

A2 - Liu, Yuling

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 16 August 2025 through 18 August 2025

ER -

Wu Y, He H, Tang Y. BEV-SI: A Lightweight and Efficient Split-Inception Framework for Multi-modal 3D Object Detection. In Li H, Wang Z, Zhao S, Sun P, Herrmann M, Zheng X, Liu Y, editors, Intelligent Vehicles - 3rd CCF Intelligent Vehicles Symposium, CIVS 2025, Revised Selected Papers. Springer Science and Business Media Deutschland GmbH. 2026. p. 160-171. (Communications in Computer and Information Science). doi: 10.1007/978-981-95-4875-0_13