Infrastructure-based perception plays a crucial role in intelligent transportation systems by providing global situational awareness and enabling cooperative autonomy. However, camera-based 3D object detection remains challenging under multi-view infrastructure setups, heterogeneous camera configurations, degraded sensing conditions, diverse road layouts, and adverse environmental conditions. We introduce MIC-BEV, a Transformer-based bird's-eye-view (BEV) perception framework for infrastructure-based multi-camera 3D object detection. MIC-BEV supports a variable number of cameras with heterogeneous intrinsic and extrinsic parameters and uses a graph-enhanced fusion module to adaptively integrate multi-view image features by modeling geometric relationships between cameras and the BEV grid. To support training and evaluation, we introduce M2I, a synthetic dataset featuring diverse camera layouts, road geometries, environmental conditions, and sensor-degradation cases. Experiments on the simulated M2I dataset and the real-world RoScenes dataset show that MIC-BEV achieves strong 3D detection performance and improved robustness under challenging conditions. Ablation studies validate the effectiveness of its key design choices, while weight analyses show that the learned fusion weights are associated with view-reliability factors and adapt to degraded camera inputs. We further report computational cost and deployment-related factors, providing additional evidence of practical feasibility and real-world deployment potential. These results indicate that MIC-BEV is a promising framework for infrastructure-based perception. The dataset and source code are available at: https://github.com/HandsomeYun/MIC-BEV.
Figures & tables
Fig. 1: Overview of the proposed MIC-BEV framework for multi-camera infrastructure perception. (a) Example scenarios of infrastructure sensing with diverse road geometry and camera configuration. (b) Multiple cameras capture the scene from different viewpoints and project objects onto a defined BEV grid. (c) MIC-BEV encodes the geometric relations between each camera and BEV cell, such as distances, angles, and height differences, to construct edge features for a graph neural network (GNN). (d) The GNN in MIC-BEV performs geometric relation-aware fusion, assigning importance weights to each camera for adaptive multi-view feature aggregation.
Fig. 2: Overview of the MIC-BEV architecture. The model takes multi-view images from a variable number of infrastructure-mounted cameras as input and extracts features through a shared backbone. A camera masking module applies random dropout or Gaussian blur to simulate degraded views. The extracted features are fused into a BEV representation via Transformer layers with temporal self-attention and our proposed Relation-Enhanced Spatial Cross-Attention . GAT networks are used to dynamically assign view-dependent weights based on camera node features and geometric relations between the camera and its visible BEV cells. The resulting BEV features are used for both object detection and BEV semantic segmentation tasks.
TABLE I: Comparison of infrastructure perception datasets. Weather and map availability are key attributes, while variation captures diversity across layouts, locations, and fields of view. Prior datasets often have fixed camera setups and narrow scene types, whereas M2I offers broader diversity and richer supervision.
Fig. 3: Scene diversity in the M2I dataset. (a) Examples of camera placement configurations across various intersection types, ranging from single-camera highway ramps to multi-camera urban intersections. (b) Diverse weather and illumination conditions, including rain, fog, dusk, and night scenes. (c) Different levels of image blur representing sensor degradation from light to heavy. These variations illustrate the broad coverage and realism of M2I for infrastructure perception.
Fig. 4: M2I Dataset Composition. (a) Scene composition showing splits by weather, lighting, road geometry, and camera count. (b) Agent composition illustrating class proportions, velocity profiles, and footprint areas across object types. (c) Camera composition showing histograms of camera height, pitch, and yaw across all scenes. These statistics demonstrate the diversity and balance of the M2I dataset for multi-camera perception.
Method
Normal
Robust
Extreme Weather
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAP ↑
NDS ↑
mATE ↓
mASE ↓
Lift-Splat-Shoot
0.446
0.486
0.742
0.489
0.336
0.416
0.781
0.510
0.367
0.340
0.764
0.631
PETR
0.596
0.713
0.301
0.107
0.415
0.521
0.595
0.156
0.440
0.536
0.689
0.193
BEVFormer
0.637
0.755
0.235
0.072
0.513
0.679
0.288
0.084
0.445
0.579
0.535
0.131
PETRv2
0.651
0.760
0.213
0.093
0.505
0.666
0.295
0.156
0.584
0.655
0.530
0.127
StreamPETR
0.677
0.778
0.211
0.067
0.512
0.678
0.293
0.110
0.591
0.678
0.489
0.102
TABLE II: Comparison of 3D object detection performance on the M2I test set under three evaluation conditions: Normal, Robust (e.g., sensor dropout or degradation), and Extreme Weather (e.g., heavy rain or fog).
Fig. 5: Qualitative 3D detection results on the M2I dataset. MIC-BEV accurately detects multiple object types across diverse urban scenes with varying traffic densities and weather or lighting conditions.
Fig. 6: Qualitative comparison between MIC-BEV and baseline models on the M2I dataset. Compared to BEVFormer and DETR3D, MIC-BEV produces more accurate and consistent detections across day, night, and extreme-weather scenes by leveraging relation-aware multi-view fusion. In the night scene, a pedestrian partially occluded by a truck is missed by the baselines but successfully detected by MIC-BEV.
Method
Easy
Hard
Average
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAOE ↓
mAP ↑
NDS ↑
mATE ↓
mASE ↓
mAOE ↓
NDS ↑
SOLOFusion
0.129
0.308
0.878
0.144
0.517
0.066
0.202
0.844
0.144
1.000
0.255
BEVDet4D
0.200
0.428
0.896
0.094
0.041
0.139
0.393
0.922
0.099
0.038
0.411
BEVDet
0.299
0.506
0.742
0.079
0.042
0.184
0.445
0.754
0.087
0.043
0.476
StreamPETR
0.513
0.619
0.690
0.102
0.032
0.284
0.496
0.739
0.107
0.031
0.558
PETRv2
0.587
0.674
0.590
0.090
0.032
0.414
0.580
0.633
0.100
0.029
0.627
TABLE III: Comparison of 3D object detection performance on the RoScenes test set under Easy and Hard levels.
Fig. 7: Qualitative results of MIC-BEV on RoScenes. Each example shows multi-camera detections and the corresponding fused BEV view over an 800-meter range. MIC-BEV maintains spatially consistent detections across diverse lighting and camera conditions. Orange-circled regions highlight cases where complementary views help recover occluded observations.
View-reliability factor
M2I-Normal
RoScenes
Projection validity
0.894
0.818
Viewing angle deviation
-0.310
-0.514
Distance to image boundary
0.757
0.728
Camera-to-object distance
0.136
-0.052
Annotated occlusion level
–
-0.138
TABLE IV: Spearman correlations between learned ReSCA weights and object-camera view-reliability factors.
Fig. 8: Examples of visual degradation in M2I and RoScenes. Orange-highlighted views indicate degraded cameras affected by blur, while other cameras provide clearer complementary observations. (a) and (b) show representative degraded-camera scenarios in M2I and RoScenes, while (c) illustrates the mild, medium, and severe blur levels used in our robustness analysis.
Condition
Weight Drop
mAP ↑
NDS ↑
Clean view
–
0.811
0.854
Mild blur ( σ=10 , kernel=11)
6.3%
0.789
0.840
Medium blur ( σ=10 , kernel=61)
12.6%
0.767
0.828
Severe blur ( σ=20 , kernel=121)
14.2%
0.748
0.817
Camera masking
20.8%
0.747
0.816
TABLE V: Relative ReSCA weight change under visual degradation on the four-camera M2I-Normal test set.
Fig. 9: Camera-wise importance weights for BEV grid cells at an intersection. For each camera (Cam1-Cam4), the perspective view and corresponding BEV importance weight map are shown. Numbered objects (①-⑤) are representative examples discussed in the text. Colored markers in the BEV weight map denote objects visible in the corresponding camera, while uncolored markers denote occluded or out-of-view objects. The result shows that each BEV cell adaptively assigns higher weights to the camera views that provide the most informative observations based on visibility and viewing geometry.
Cam. Masking
BEV Seg.
ReSCA
M2I-Normal
M2I-Robust
mAP ↑
NDS ↑
mAP ↑
NDS ↑
✗
✗
✗
0.637
0.755
0.513
0.679
✓
✗
✗
0.651
0.767
0.624
0.701
✓
✓
✗
0.717
0.792
0.669
0.737
✓
✓
✓
0.793
0.844
0.734
0.809
TABLE VI: Ablation study of MIC-BEV components on M2I. Cam. Masking : random camera masking; BEV Seg. : the auxiliary BEV segmentation head; ReSCA : Relation-enhanced spatial cross-attention module
Cam. Feature
Distances
Angles
M2I-Normal
mAP ↑
NDS ↑
✗
✗
✗
0.717
0.792
✗
✓
✓
0.738
0.794
✓
✗
✗
0.733
0.789
✓
✓
✗
0.756
0.806
✓
✓
✓
0.793
0.844
TABLE VII: Influence of incorporating camera features, distance, and angles in the GNN-based multi-view fusion module.
Fig. 10: Representative failure cases of MIC-BEV. ReSCA weight maps and corresponding camera views illustrate two representative failure modes: suboptimal view weighting, where useful local evidence is underweighted (M2I), and information-limited detection, where all available views provide insufficient evidence despite reasonable weighting (RoScenes).
Category
Condition
M2I-Normal
RoScenes
Base
Aug.
Base
Aug.
Clean
Clear view
0.850
0.844
0.745
0.749
Visual degradation
Mild blur
0.789
0.829
0.731
0.744
Medium blur
0.773
0.814
0.710
0.735
Severe blur
0.772
0.797
0.709
0.724
Camera masking
0.769
0.789
0.706
0.717
TABLE VIII: Robustness evaluation under visual degradation, camera masking, and calibration noise on M2I-Normal and RoScenes. Base denotes training on clean data, while Aug. denotes training with degradation data augmentation. Results are reported using NDS.
Method
Params (M)
Memory (GB)
FPS
mAP ↑
LSS
21.51
2.05
17.02
0.446
PETR
37.26
2.77
16.51
0.596
BEVFormer
68.67
3.64
9.97
0.637
PETRv2
61.53
3.28
12.94
0.651
StreamPETR
37.26
2.65
14.05
0.677
BEVNeXt
65.43
2.36
7.66
0.681
TABLE IX: Performance and inference cost of detection models on a single NVIDIA L40S GPU with batch size 1.
Category
Setting
Params
Memory
FPS
mAP ↑
NDS ↑
BEV Resolution
100 × 100
65.793
2.59
14.24
0.765
0.815
200 × 200 †
73.498
3.68
8.82
0.793
0.844
300 × 300
86.324
7.62
3.02
0.818
0.859
Camera Count
1 cam
73.494
2.76
15.09
0.451
0.623
2 cams
73.496
3.05
12.92
0.625
0.727
3 cams
73.497
3.43
10.10
0.726
0.790
TABLE X: Scalability analysis on M2I. The BEV-resolution experiment fixes the max input to 4 cam, while the camera-count experiment fixes the BEV resolution at 200×200 and varies the max camera capacity. † denotes the default setting.
Monocular 3D object detection methods increasingly rely on depth-guided transformers, but they operate entirely in the perspective image plane where spatial relationships are distorted by foreshortening. We propose MonoMIP, a dual-decoder framework that adds complementary 3D geometric reasoning through Bird's-Eye View (BEV) features constructed via Multi-Height Inverse Perspective Mapping. We perform IPM at K elevation planes spanning ground to rooftop, producing an ordered stack of height-conditioned metric BEV features that reduces above-ground misprojection and exposes depth-dependent appearance variation across planes. A BEV decoder with multi-scale deformable attention aggregates this multi-height evidence per query, while a depth-guided 3D decoder provides complementary perspective-space predictions. Without extra training data, MonoMIP attains 25.84% Moderate AP3D|R40 on KITTI val and 21.66% on the official test set for Car, outperforming recent published monocular methods on both the KITTI and Waymo benchmarks.
Simultaneous perception of 2D objects in perspective view and 3D objects in Bird's Eye View (BEV) is challenging for multi-camera autonomous driving. Existing two-stage pipelines use 2D results only as a one-time cue for 3D detection. We propose SimPB++, which simultaneously detects 2D objects in perspective and 3D objects in BEV from multiple cameras. It unifies both tasks into an end-to-end model with a hybrid decoder architecture, coupling multi-view 2D and 3D decoders interactively. Two novel modules enable deep interaction: Dynamic Query Allocation adaptively assigns 2D queries to 3D candidates, and Adaptive Query Aggregation refines 3D representations using multi-view 2D features, forming a cyclic 3D-2D-3D refinement. For multi-view 2D detection, we use Query-group Attention for intra-group communication. We also design a Crop-and-Scale strategy for long-range perception and a Propagating Denoising strategy with an auxiliary RoI detector. SimPB++ supports mixed supervision with 2D-only and fully annotated data, reducing reliance on expensive 3D labels. Experiments show state-of-the-art performance on nuScenes for both tasks and strong long-range detection (up to 150m) on Argoverse2.
Radar-camera BEV perception often suffers from degraded performance when evaluated across datasets, as changes in driving scenes, sensor configurations, and environmental conditions can alter both the input observations and the internal fused representations. This work studies this issue from the perspective of source-domain variation modeling, aiming to improve the robustness of BEV-based 3D detectors without relying on target-domain samples. We introduce a framework that characterizes visual scene variations in the frequency domain and uses them to synthesize diverse source-domain views. By comparing the resulting fused BEV representations, the framework further captures how image-level variations influence multi-modal BEV features. These variation patterns are then used to regularize the detector, encouraging the learned fusion space to remain stable under latent scene changes. The proposed method is applied only during training and leaves the inference pipeline unchanged. Experiments on cross-dataset radar-camera 3D detection between View-of-Delft and TJ4DRadSet demonstrate consistent improvements over multiple BEV fusion backbones, and the gains remain effective when a small amount of target-domain data is available.