ESAFusion: LiDAR--4-D Radar Fusion via Local Geometric Complementation and Multiscale Adaptive Interaction for 3-D Object Detection

Gang Ma; Senjie Hu; Junjie Liu; Chao Wang; Hui Wei

cs.CVSep 13, 2026

ESAFusion: LiDAR--4-D Radar Fusion via Local Geometric Complementation and Multiscale Adaptive Interaction for 3-D Object Detection

Authors: Gang Ma, Senjie Hu, Junjie Liu, Chao Wang, Hui Wei

Organizations: School of Future Technology, Shanghai University, Shanghai 200444, China

Abstract

LiDAR--4-D radar fusion combines accurate spatial geometry with motion and reflectivity cues from radar, offering a promising solution for 3-D object detection in complex driving environments. However, sparse radar observations and differences in spatial sampling between the two modalities complicate reliable cross-modal complementation. Moreover, the relative importance of modalities and feature scales varies across spatial regions, making adaptive fusion challenging. To address these challenges, we propose ESAFusion, an evidence-aware and scale-adaptive framework that combines local geometric complementation with multiscale adaptive interaction. Specifically, we introduce an Evidence-Aware Radar Selection (ERS) module to suppress radar clutter using motion and observation-quality evidence while retaining foreground confidence for subsequent fusion. Then, the Pillar-Level Complementary Encoder (PCE) improves cross-modal complementation under mismatched spatial sampling using local geometric support from neighboring LiDAR pillars. We further design an Intra- and Inter-Scale Adaptive Fusion (ISAF) module to adaptively adjust the contributions of different modalities and feature scales in bird's-eye-view (BEV) space. Extensive experiments on the View-of-Delft (VoD) dataset show that ESAFusion achieves the highest mean average precision (mAP) among the compared methods, reaching 74.60% in the Entire Annotated Area and 88.89% in the Driving Corridor. It also attains the highest average precision (AP) for Cyclist among these methods in both regions while running at 19.23 FPS. Evaluations on VoD-Fog further demonstrate robustness under progressively degraded LiDAR observations. The source code will be made publicly available at https://github.com/SenJieHu549/ESAFusion.

Figures & tables

Fig. 1: Comparison between strictly co-located fusion and ESAFusion. (a) Strict co-location leaves radar pillars without LiDAR support when same-cell LiDAR observations are absent. (b) ESAFusion complements such radar pillars using neighboring LiDAR observations within a local $3\times 3$ BEV neighborhood.

Fig. 2: Quantitative analysis on the VoD validation set. (a) Class-wise availability of neighboring LiDAR observations. (b) Local LiDAR support for false-positive and true-positive detections. (c) Radar motion evidence for false-positive and true-positive detections.

Fig. 3: Overview of the ESAFusion framework. (a) Evidence-Aware Radar Selection (ERS) models motion and observation-quality evidence to refine radar attributes and suppress low-confidence clutter. (b) Pillar-Level Complementary Encoder (PCE) performs co-located attribute complementation, incorporates geometrically consistent local LiDAR support, and enhances radar pillars with compact radar evidence. (c) Intra- and Inter-Scale Adaptive Fusion (ISAF) adaptively gates LiDAR and radar features within each scale and performs support-conditioned inter-scale recalibration to produce a fused BEV representation for detection.

Fig. 4: Architecture of the PCE module, comprising Co-Located Multimodal Complementation (CMC), Asymmetric Geometric Injection (AGI), and Compact Radar Evidence Enhancement (CREE).

Fig. 5: Architecture of the ISAF module, comprising Evidence-Refined Intra-Scale Modality Gating and Support-Conditioned Inter-Scale Recalibration.

Methods	Modality	Entire Annotated Area (EAA)				Driving Corridor (RoI)
Methods	Modality	Car	Ped.	Cyc.	mAP	Car	Ped.	Cyc.	mAP
PointPillars (CVPR’19) [ 20 ]	R	38.89	32.07	65.02	45.32	71.07	42.61	86.61	66.77
SMURF (T-IV’24) [ 11 ]	R	42.31	39.09	71.50	50.97	71.74	50.54	86.87	69.72
4DRadDet (ICRA’25) [ 37 ]	R	42.03	40.70	71.61	51.44	72.12	51.18	87.95	70.42
MAFF-Net (RA-L’25) [ 12 ]	R	42.33	46.75	74.72	54.59	72.28	57.81	87.40	72.50
RadarGaussianDet3D (RA-L’26) [ 38 ]	R	40.70	42.40	73.00	52.03	71.20	51.70	89.00	70.63

TABLE I: Comparative AP (%) results on the VoD validation set. The best results are bold, and the second best are underlined. R, C, and L denote 4-D radar, camera, and LiDAR, respectively

Methods	Modality	3-D AP				BEV AP
Methods	Modality	Easy	Moderate	Hard	Average	Easy	Moderate	Hard	Average
PointPillars (CVPR’19) [ 20 ]	R	21.58	15.21	14.15	16.98	39.52	27.85	25.75	31.04
PV-RCNN (CVPR’20) [ 21 ]	R	18.37	13.04	11.95	14.45	38.82	27.37	25.54	30.58
PV-RCNN++ (IJCV’23) [ 46 ]	R	18.20	13.05	12.04	14.43	31.73	23.65	22.24	25.87
Voxel Mamba (NeurIPS’24) [ 44 ]	R	20.63	14.70	13.89	16.41	36.27	27.66	25.87	29.93
PointPillars (CVPR’19) [ 20 ]	L	47.30	35.05	32.75	38.37	52.76	39.79	37.40	43.32

TABLE II: Comparative 3-D and BEV AP (%) results for the Car class on the Astyx HiRes2019 validation set at an IoU threshold of 0.5. The best results are bold, and the second best are underlined. R and L denote 4-D radar and LiDAR, respectively

Fig. 6: Qualitative comparison of L4DR, SVEFusion, and the proposed ESAFusion on the VoD validation set. Green, yellow, and black boxes denote pedestrians, cyclists, and cars, respectively. Orange and red circles indicate false negatives and false positives, respectively.

Fig. 7: Distance-wise cross-modal support availability on the VoD validation set. (a) Percentage of radar pillars without LiDAR observations in the same BEV cell. (b) Among these missing-support pillars, the percentage having at least one LiDAR-supported cell within the local

3\times 3

neighborhood. Here,

n

denotes the number of valid validation frames in each distance interval.

Distance	Method	Car AP	Ped. AP	Cyc. AP	mAP
0–20 m	L4DR [ 6 ]	81.44	75.31	83.99	80.25
	SVEFusion [ 3 ]	81.73	77.61	85.53	81.62
	ESAFusion	81.68	76.62	86.04	81.45
20–40 m	L4DR [ 6 ]	71.21	57.56	88.49	72.42
	SVEFusion [ 3 ]	72.54	65.56	87.87	75.32
	ESAFusion	72.10	67.56	88.67	76.11

TABLE III: Distance-wise AP (%) results on the VoD validation set. The best and second-best results within each distance interval are highlighted in bold and underlined, respectively

Fig. 8: Distance-wise qualitative comparison on the VoD validation set. (a) ESAFusion detects a Cyclist at 48.31 m that is missed by L4DR and SVEFusion. (b) ESAFusion correctly classifies a Pedestrian at 27.71 m that is misclassified as a Cyclist by both baselines.

Fog Level	Methods	Modality	Car (IoU $=0.5$ )			Pedestrian (IoU $=0.25$ )			Cyclist (IoU $=0.25$ )
Fog Level	Methods	Modality	Easy	Mod.	Hard	Easy	Mod.	Hard	Easy	Mod.	Hard
0 (W/o Fog)	PointPillars [ 20 ]	L	84.90	73.50	67.50	62.70	58.40	53.40	85.50	79.00	72.70
	L4DR [ 6 ]	L+R	85.00	76.60	69.40	74.40	72.30	65.70	93.40	90.40	83.00
	SVEFusion [ 3 ]	L+R	87.38	76.69	69.04	78.91	74.39	67.59	93.14	89.24	81.93
	ESAFusion	L+R	86.82	76.87	69.84	78.69	74.43	68.27	95.26	90.45	83.35
1	PointPillars [ 20 ]	L	79.90	72.70	67.00	59.90	55.60	50.50	85.50	78.20	72.00

TABLE IV: Quantitative results of different methods on the VoD-Fog dataset using KITTI metrics under various fog levels. The best results are bold

Fig. 9: Qualitative 3-D object detection results on the VoD-Fog validation set under different simulated fog levels. Gray point clouds represent LiDAR data, while orange points represent 4-D radar data. Red boxes indicate ground-truth objects, and blue boxes indicate predicted bounding boxes.

Module			EAA				RoI
ERS	PCE	ISAF	Car	Ped.	Cyc.	mAP	Car	Ped.	Cyc.	mAP
–	–	–	66.20	55.99	75.87	66.02	88.70	69.40	88.30	82.13
✓	–	–	66.72	58.80	78.10	67.87	89.55	72.89	89.07	83.84
–	✓	–	69.60	67.18	80.47	72.41	90.78	76.13	90.08	85.66
✓	✓	–	70.25	68.54	81.92	73.57	90.90	77.83	90.48	86.40
–	–	✓	69.40	67.55	81.80	72.92	90.75	76.95	93.75	87.15

TABLE V: Ablation of different modules on the VoD validation set

CMC	AGI	CREE	EAA mAP	RoI mAP
✓	–	–	73.02	86.71
✓	–	✓	73.55	87.28
✓	✓	–	74.01	88.35
✓	✓	✓	74.60	88.89

TABLE VI: Ablation study of the PCE module on the VoD validation set

Range	EAA mAP	RoI mAP
off	73.55	87.28
4-connected	74.12	88.52
$3\times 3$	74.60	88.89
$5\times 5$	73.90	87.84

TABLE VII: Ablation study of the neighborhood settings for AGI on the VoD validation set

Setting	EAA mAP	RoI mAP
w/o ISAF	73.57	86.40
w/o Intra-Scale Gating	72.16	86.01
w/o Inter-Scale Recalibration	74.20	87.64
Full ISAF	74.60	88.89

TABLE VIII: Ablation study of the ISAF module on the VoD validation set

ERS	PCE	ISAF	GFLOPs	Latency (ms)	Params (M)
–	–	–	88.19	13.55	18.05
$\checkmark$	–	–	91.85 (+3.66)	18.42 (+4.87)	20.94 (+2.89)
$\checkmark$	$\checkmark$	–	91.90 (+0.05)	43.34 (+24.92)	21.16 (+0.22)
$\checkmark$	$\checkmark$	$\checkmark$	255.67 (+163.77)	52.01 (+8.67)	62.07 (+40.91)

TABLE IX: Ablation study of GFLOPs, latency, and parameter count for different module configurations

Explore similar work

Sep 24, 2026cs.CV

SARFusion: Scene-Aware Routing Fusion for Robust Camera-LiDAR 3D Object Detection

Camera-LiDAR fusion has become a prevailing paradigm for 3D object detection in autonomous driving. However, existing fusion detectors often establish strong inter-modality dependencies by decoding object queries from tightly coupled multimodal representations. Under corrupted driving conditions, such dependencies make the detector vulnerable to unreliable modalities, where degraded observations may interfere with reliable modality-specific evidence and lead to suboptimal predictions. Moreover, modality reliability can vary across both global driving scenes and individual object queries, requiring adaptive fusion decisions at a finer granularity. To bridge this gap, we reformulate robust camera-LiDAR fusion as a scene-aware branch routing problem and propose SARFusion, a robust 3D object detector. Instead of producing detections from a single fused representation, SARFusion decouples object-query decoding into three parallel reasoning branches: a camera branch, a LiDAR branch, and a camera-LiDAR fusion branch. Guided by a Scene Reliability Prior estimated from the global driving context, SARFusion further incorporates object-level evidence to route each query to the most suitable branch. This query-wise routing strategy alleviates harmful cross-modal interference while preserving the benefits of multimodal fusion when complementary cues are trustworthy. On the nuScenes test set, SARFusion achieves strong performance with 72.5 mAP and 74.4 NDS. Extensive analyses demonstrate its robustness under challenging conditions, including sensor corruptions and environmental changes.

Yuting Zhao, Ziyi Zheng, Shuxiao Li

Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · Wuhan College, Wuhan, China

3D Object Detection Multi-Sensor Fusion Lidar Autonomous Driving 3D Generation

Ask PDF

Jul 5, 2026cs.CV

Sparse4D-Radar: An Efficient and Robust Framework for Surround-View 3D Object Detection via 4D Radar-Camera Fusion

In recent years, 4D imaging radar has gained wide attention in autonomous driving for its robustness against harsh weather and ability to output target velocity. Nevertheless, mainstream 4D radar-camera fusion methods only support front-view perception, lacking mature solutions for surround-view sensing. Directly expanding these pipelines to full 360° coverage introduces excessive computation cost and limits real-world deployment. To tackle these limitations, this work proposes Sparse4D-Radar, an efficient robust surround-view multi-modal fusion framework. We first design a Deformable Fusion module to embed radar-camera features into sparse queries, constructing the lightweight base version Sparse4D-Radar-Base. Two dedicated modules are further introduced to boost localization accuracy and modality stability: Velocity-Consistency Sampling (VCS) refines features via radar velocity cues for motion awareness, and Adaptive Modality Gating (AMG) dynamically adjusts cross-modal fusion weights according to feature confidence. Combining all components, we build Sparse4D-Radar-Acc for high-precision detection demands. Comprehensive experiments on OmniHD-Scenes verify that our approach achieves state-of-the-art surround-view 3D detection performance. Compared with prior arts, our method obtains over 7% mAP and 10% ODS improvements under complex driving scenes while running at nearly 10 FPS, striking a favorable trade-off among detection accuracy, environmental robustness and inference efficiency. Our open-source code is available at https://github.com/Aiuan/Sparse4D-Radar.

Fuyuan Ai, Yuchen Tan, Jiehui Chen +2

State Key Laboratory of Ocean Sensing and Ocean College, Zhejiang University, Zhoushan 316021, China, and also with the Engineering Research Center of Oceanic Sensing Technology and Equipment, Ministry of Education, Zhejiang University, Zhoushan 316021, China · Donghai Laboratory, Zhoushan 316021, China, also with the Ocean College, Institute of Marine Electronics and Intelligent Systems, Zhejiang University, Zhoushan 316021, China, and also with the Engineering Research Center of Oceanic Sensing Technology and Equipment, Ministry of Education, Zhejiang University, Zhoushan 316021, China

4D Radar 3D Object Detection Multi-Sensor Fusion Autonomous Driving 3D Generation

Ask PDF

Jul 6, 2026cs.CV

RAF: Reliability-Aware Fusion of Camera, LiDAR, and 4D RADAR for Robust 3D Object Detection in Adverse Weather

Robust 3D object detection in adverse weather conditions is challenging due to sensor limitations. Although combining complementary modalities such as LiDAR and 4D RADAR has shown promise, the sparsity of these sensors becomes apparent in adverse weather with reduced reflections, leading to objects with few or no point cloud returns. To address this limitation, camera sensors provide visual cues even when LiDAR and RADAR signals are weakened. However, cameras themselves are also vulnerable to adverse weather, where some regions become unreliable due to snow or rain occluding the camera lens. While some camera-fusion methods designed for adverse weather learn to weigh image regions via confidence maps, these maps receive no direct supervision and are learned solely through the detection loss. We introduce Reliability-Aware Fusion (RAF), which explicitly supervises per-pixel reliability estimation and provides a direct learning signal for identifying and suppressing unreliable visual cues. Our framework leverages pretrained LiDAR-RADAR networks, keeping their backbones frozen while only training the added camera branch, BEV fusion encoder, and detection head. Extensive experiments on the K-Radar and VoD datasets demonstrate that integrating RAF consistently improves detection accuracy over LiDAR-RADAR baselines, achieving up to +6.5

AP_{BEV}

and +7.4

AP_{3D}

gains. Code is available at https://github.com/parkie0517/RAF.

Heejun Park, Jaeseok Jeong, Kuk-Jin Yoon

Visual Intelligence Lab., KAIST

3D Object Detection 4D Radar Lidar Uncertainty-Aware Fusion Camera

Ask PDF