Post Fusion Bird's Eye View Feature Stabilization for Robust Multimodal 3D Detection

Trung Tien Dong; Dev Thakkar; Arman Sargolzaei; Xiaomin Lin

cs.CVMar 5, 2026

Post Fusion Bird's Eye View Feature Stabilization for Robust Multimodal 3D Detection

Authors: Trung Tien Dong, Dev Thakkar, Arman Sargolzaei, Xiaomin Lin

Organizations: College of Engineering, University of South Florida, Tampa, FL 33620, USA

Abstract

Camera-LiDAR fusion is widely used in autonomous driving to enable accurate 3D object detection. However, bird's-eye view (BEV) fusion detectors can degrade significantly under domain shift and sensor failures, limiting reliability in real-world deployment. Existing robustness approaches often require modifying the fusion architecture or retraining specialized models, making them difficult to integrate into already deployed systems. We propose a Post Fusion Stabilizer (PFS), a lightweight module that operates on intermediate BEV representations of existing detectors and produces a refined feature map for the original detection head. The design stabilizes feature statistics under domain shift, suppresses spatial regions affected by sensor degradation, and adaptively restores weakened cues through residual correction. Designed as a near-identity transformation, PFS preserves performance while improving robustness under diverse camera and LiDAR corruptions. Evaluations on the nuScenes benchmark demonstrate that PFS achieves state-of-the-art results in several failure modes, notably improving camera dropout robustness by +1.2% and low-light performance by +4.4% mAP while maintaining a lightweight footprint of only 3.3 M parameters.

Figures & tables

Figure 1: PFS inserts a lightweight correction module between the BEV fusion stage and the detection head of an existing detector.

Figure 2: Overall PFS architecture. A lightweight Feature Stabilizer is placed between the fused BEV feature map and the frozen detection head. Operating on $\mathbf{F}\in\mathbb{R}^{C\times H\times W}$ , it applies (1) channel shift normalization, (2) spatial reliability suppression, and (3) degradation aware gated residual correction with semantic and geometric experts to produce $\tilde{\mathbf{F}}$ , which is then fed to the unchanged head. All stabilizer blocks are identity initialized for safe deployment.

Figure 3: LiDAR and Camera corruptions at each severity levels.

Sensor	Corruption Type	L1	L2	L3
Camera	Rain Lens ( $\alpha$ / drops)	0.15 / 80	0.45 / 200	0.70 / 350
	Low Light ( $\gamma$ / $\sigma$ )	1.8 / 0.018	3.2 / 0.045	5.5 / 0.10
	Occlusion (Area $\mathcal{C}$ )	15%	35%	60%
	Dropout (Views $N$ )	1	3	6
LiDAR	Beam Red. (Step $k$ )	2 (16-ch)	4 (8-ch)	8 (4-ch)
	Range Drop ( $P_{max}@r_{max}$ )	0.70 @ 50m	0.85 @ 45m	0.95 @ 40m

Table I : Sensor Corruption Severity Levels. Parameters define the physical intensity of simulated degradation across three severity levels.

			LiDAR corruptions								Camera corruptions
Method	Clean		Beam Reduction		Range Dropout		Sector Dropout		Miscalibration		Rain Lens		Low Light		Occlusion		Dropout
Method			8-beams		95% @ 40m		90 ∘ Sector		2.0 ∘ , 0.2m		350 drops		$\gamma=5.5$		60% Area		6 drops
	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS
BEVFusion † [ 11 ]	69.2	59.7	12.1	24.8	52.6	50.0	54.7	52.4	46.8	44.1	60.0	55.0	58.8	54.4	64.6	57.3	62.6	56.3
UniBEV † [ 18 ]	65.3	57.2	25.6	29.9	53.2	50.0	56.5	52.2	48.0	45.1	56.5	52.8	55.4	52.3	61.5	55.5	56.7	53.1
CMT † [ 19 ]	70.1	72.9	32.6	45.7	58.5	65.2	60.7	67.0	50.1	58.0	58.9	66.6	55.4	64.3	64.1	70.2	62.0	68.6

Table II : Comparison under various sensor corruptions and sensor failures on nuScenes val dataset. ’†’ indicates reproduced results using their open-source code. Italic denotes the degree of sensor failure. All sensor corruptions severity are level 3. PFS achieves state-of-the-art performance across most tasks ( BEST / SECOND BEST ).

Method	Modality	Fog	Snow	Sunlight
PGD [ 21 ]	C	12.8	2.3	22.8
FCOS3D [ 22 ]	C	13.5	2.0	17.2
DETR3D [ 23 ]	C	27.9	5.1	41.6
BEVFormer [ 6 ]	C	32.8	5.7	41.7
PointPillars [ 24 ]	L	24.5	27.6	23.7
SSN [ 25 ]	L	41.6	46.4	40.3

Table III : Comparison under extreme weather conditions on nuScenes-C. ‘L’ and ‘C’ represent camera, LiDAR, respectively ( BEST / SECOND BEST ).

Figure 4: Real time camera and LiDAR synchronization during deployment. Each card shows a timestamp aligned RGB frame (top) and raw LiDAR point cloud (bottom) in the ego frame.

Method	Clean		Rain Lens		Low Light		Beam Red.		Miscalib.
Method	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS
A0: BEVFusion [ 11 ]	69.2	59.7	60.0	55.0	58.8	54.4	12.1	24.8	46.8	44.1
A1: + Block 1	68.7	60.0	61.6	56.0	61.4	56.7	11.9	25.5	46.8	44.6
A2: + Blocks 1 and 2	68.9	60.0	62.3	56.5	61.8	57.0	14.9	25.0	47.0	44.6
A3: + All blocks	69.5	71.2	63.6	68.3	63.2	64.9	17.5	27.1	47.5	55.5
A0: UniBEV [ 18 ]	65.3	57.2	56.5	52.8	55.4	52.3	25.6	29.9	48.0	45.1

Table IV : Block ablation across detectors on nuScenes val. Each detector is evaluated with progressively enabled blocks: A0 (base), A1 (+Block 1), A2 (+Blocks 1&2), A3 (+All blocks).

Figure 5: Real-world data collection and deployment platform

Method	mAP		Efficiency / Overhead
Method	Day	Night	Params (M)	Latency (ms)	FPS	Overhead (%)
BEVFusion [ 11 ]	32.87	24.08	40.8	70.9	14.1	-
+ PFS (Proposed)	35.33	29.20	44.1	76.6	13.1	8.1%
UniBEV [ 18 ]	33.87	24.75	72.7	417.0	2.4	-
+ PFS (Proposed)	36.65	28.83	76.0	436.0	2.3	4.5%

Table V : Performance and efficiency analysis on real-world physical data. PFS improves mAP under both day-time and night-time conditions. While the additional 3.3 M parameters slightly increase latency, the module remains a lightweight add-on.

Multimodal Robustness

Multimodal robustness studies how systems combining different data types remain reliable when inputs are missing, noisy, shifted, or misleading.

See all 279 papers

SB-BEVFusion: Enhancing the Robustness against Sensor Malfunction and Corruptions
May 12, 2026Markus Essl, Marta Moscati, Mubashir Noman +4Multimodal Robustness Camera-LiDAR Fusion
Johannes Kepler University Linz, Austria · MBZUAI, UAE · Macquarie University, Sydney, Australia +1
Geometry-Aware Fisheye-LiDAR Fusion for Robust 3D Object Detection in Low-Overlap Setups
Jun 7, 2026Xiangzhong Liu, Xihao Wang, Hao ShenCamera-LiDAR Fusion Bird's-Eye-View Representation
Technical University of Munich, 80333 Munich, Germany
Can BEV Perception Gracefully Degrade under Sensor Failures?
May 29, 2026Haifa Zhang, Yijing Wang, Haoyu Wang +2Missing-Modality Learning Multimodal Robustness
Tianjin Key Laboratory of Intelligent Unmanned Swarm Technology and System, School of Electrical and Information Engineering, Tianjin University, Tianjin, China · Key Laboratory of System Control and Information Processing, Ministry of Education of China, Shanghai, China

May 12, 2026J/K move · Enter open · S save

SB-BEVFusion: Enhancing the Robustness against Sensor Malfunction and Corruptions

Markus Essl, Marta Moscati, Mubashir Noman, Muhammad Zaigham Zaheer+3

Johannes Kepler University Linz, Austria · MBZUAI, UAE · Macquarie University, Sydney, Australia+1

Multimodal Robustness Camera-LiDAR Fusion Autonomous Driving 3D Object Detection

Open paper AskPDF arXiv

Multimodal sensor fusion has demonstrated remarkable performance improvements over unimodal approaches in 3D object detection for autonomous vehicles. Typically, existing methods transform multimodal data from independent sensors, such as camera and LiDAR, into a unified bird's-eye view (BEV) representation for fusion. Although effective in ideal conditions, this strategy suffers from substantial performance deterioration when camera or LiDAR data are missing, corrupted, or noisy. To address this vulnerability, we develop a framework-agnostic fusion module for camera and LiDAR data that allows for handling cases when one of the two modalities is missing or corrupted. To demonstrate the effectiveness of our module, we instantiate it in BEVFusion [1], a well-established framework to combine camera and LiDAR data for 3D object detection. By means of quantitative experiments on the MultiCorrupt dataset, we demonstrate that our module achieves favorable performance improvements under scenarios of missing and corrupted modalities, substantially outperforming existing unified representation approaches across a wide range of sensor deterioration scenarios and reaching state-of-the-art performance in scenarios of corrupted modality due to extreme weather conditions and sensor failure.

Abstract

Multimodal Robustness

Explore similar work

SB-BEVFusion: Enhancing the Robustness against Sensor Malfunction and Corruptions

Geometry-Aware Fisheye-LiDAR Fusion for Robust 3D Object Detection in Low-Overlap Setups

Can BEV Perception Gracefully Degrade under Sensor Failures?